Drop psqlpy-python forks; resolve all deps through crates.io

[Dev-iL]

Summary

• Removes the 7-fork constellation (rust-postgres, rust-postgres-array, rust-postgres-interval, pgvector-rust, rust-decimal, deadpool, pyo3-async-runtimes) from Cargo.toml and rewires every fork-only call site against in-tree Rust idioms. Cargo.toml and Cargo.lock now resolve every dep through crates.io.
• New PsqlpyManager (src/driver/psqlpy_manager.rs) implements deadpool::managed::Manager with a libpq-faithful plaintext-first / TLS-fallback pair for SslMode::Allow; is_ssl_required_rejection walks the source() chain matching INVALID_AUTHORIZATION_SPECIFICATION + "no encryption" to decide when to switch sides. Non-Allow modes wrap a single inner deadpool_postgres::Manager — identical behavior to the pre-change direct path.
• New RawBuf<'a> FromSql newtype (src/value_converter/raw_buf.rs) replaces the fork-only Row::col_buffer accessor via row.try_get::<usize, RawBuf>(i).
• DEALLOCATE PREPARE calls dropped; rely on tokio-postgres Drop for StatementInner → Close('S', name) + Sync for autoclose. The Connection::drop_prepared trait method is gone.
• BinaryCopyInWriter empty-buffer trio replaced by direct CopyInSink<Bytes>::send + finish (Python side already builds the binary-format payload, so the wire framing layer was redundant).
• SslMode::VerifyCa / VerifyFull verification moved into build_tls at the postgres-openssl SslConnector layer: SslVerifyMode::PEER on the builder, plus a per-connection set_verify_hostname(false) callback for VerifyCa (the VerifyFull branch keeps openssl's default hostname-on behavior). to_internal() collapses to the three upstream-supported variants (Disable / Prefer / Require).
• postgres_array::Array::from_parts_no_panic replaced by pre-validation + upstream Array::from_parts at the single call site in src/value_converter/from_python.rs. The pre-validation reproduces the fork's empty-empty acceptance case while still using the panicking upstream constructor only after the size check passes.
• pyo3 ecosystem bumped 0.25 → 0.26 (Python 3.10 still in scope; 0.27 drops 3.10 support). Changes that would only land cleanly on pyo3 0.28+ are reverted with TODO(python-3.10-drop) markers next to each — covering Python::attach, FromPyObject<'a, 'py>::extract, OnceLockExt, the root-level PyObject rename, the FromPyObjectBound → FromPyObject<'a, 'py> move, the PyClassGuardError conversion, and the pyo3-async-runtimes::future_into_py T: Send + 'static tightening.

Why now / motivation

The forks under psqlpy-python/ had drifted from upstream — the deepest (rust-postgres) was 202 commits behind — and the maintenance cost was paid every time we tried to track an upstream security fix or feature. The actual delta we needed turned out to be five pub exposures and three convenience methods in rust-postgres, plus one workaround in rust-postgres-array. All of them now live in psqlpy itself, in well-named in-tree Rust idioms that are easier to reason about than out-of-tree forks. Six of the seven forks were Cargo.toml-only repointers; once the rust-postgres delta is gone, they collapse straight to crates.io.

Test plan

Verified locally

• cargo build --release --all-features — green
• cargo clippy -p psqlpy --all-features -- -W clippy::all -W clippy::pedantic -D warnings — clean (INV-G7)
• cargo fmt -- --check --config use_try_shorthand=true,imports_granularity=Crate — clean (INV-G6)
• Pre-commit hooks (mypy, trim trailing whitespace, rust fmt, rust clippy, rust cargo check) — all passed under git commit
• grep -nE 'git\s*=\s*"https?://github.com/psqlpy-python' Cargo.toml Cargo.lock — 0 hits (INV-G1)
• git grep -nE 'SummitSG-LLC|psqlpy-python/(rust-postgres|deadpool|pyo3-async-runtimes|rust-decimal|rust-postgres-array|rust-postgres-interval|pgvector)' — 0 hits (INV-G2)
• maturin develop --release succeeds; module imports; all six SslMode variants exposed
• Full pytest suite against a live pg14: 247 cases pass on py3.12. test_binary_copy (the path that exercises the new CopyInSink::send + finish bypass) passes. test_ssl_mode was skipped — the local postgres doesn't have TLS enabled, no POSTGRES_CERT_FILE.

Deferred to CI / follow-up

• tox -v -c tox.ini matrix across py3.10/3.11/3.12/3.13/3.14 × pg14/15/16/17 (INV-G8). The fmt and clippy gates already passed locally; the gap is multi-Python + multi-pg-version coverage.
• test_ssl_mode cert-backed run against pg14/15/16/17. Exercises the new build_tls PEER + hostname-callback split for VerifyCa vs VerifyFull, plus the new PsqlpyManager Allow-mode plaintext-first wiring.
• New libpq-Allow e2e pytest (manifest INV-G3) — needs hostssl-only postgres infra. Not added in this PR.
• New statement-cache-eviction pytest (manifest INV-G4) — not added in this PR.
• Per-AC review-agent gates (G9–G16: code-bugs, change-intent, contracts, type-safety, maintainability, ops, test-quality, prose, context-file) — left for reviewers.

Risk flags worth a second look

• R-1 statement-cache autoclose. The DEALLOCATE PREPARE removal banks on tokio-postgres dropping the last Arc<StatementInner> clone. Any path that holds a clone of Statement past the consumer's Result would silently leak the prepared statement server-side. INV-G4 (the eviction-leak test) would catch this; it's deferred.
• R-2 hostname toggle API. postgres-openssl 0.5.3 exposes MakeTlsConnector::set_callback(|cfg, _| cfg.set_verify_hostname(false)) — that's what the VerifyCa branch uses. If the openssl ConnectConfiguration::set_verify_hostname shape ever changes, this is the line that breaks (src/driver/utils.rs::build_tls).
• R-3 "no encryption" wording. is_ssl_required_rejection substring-matches "no encryption" in the postgres SQLSTATE 28000 error message. Empirically present on pg 12–17 in auth.c:ClientAuthentication, but if a pg version differs, the libpq-Allow retry path silently regresses to never-retry. INV-G3 across pg14–17 would catch this; deferred.
• R-5 RawBuf hot-path cost. row.try_get::<usize, RawBuf>(i) adds a FromSql dispatch hop relative to the direct col_buffer(i) field read. The trait impl returns the slice unchanged, so the optimizer should inline through it, but no benchmark was run.

Notes for the reviewer

• The [patch.crates-io] workaround mentioned in PG-6 of the manifest didn't end up being needed in the final landing form — once every direct dep moved to crates.io together with the source-side fork-API removals, the postgres-types / FromSql trait-unification issue disappeared on its own.
• The TODO(python-3.10-drop) markers are intentionally co-located with their reverted form so the next pyo3 bump becomes a single grep-and-replace pass rather than a re-derivation.
• Connection::drop_prepared was removed from the trait, not deprecated. If any out-of-tree consumer relied on this method, that's a breaking change at the Rust API boundary (the Python API is unaffected).