feat(handlers): add built-in double-click plot handlers for common types

[matt-pharr]

Summary

Adds built-in double-click plot handlers so that double-clicking a tree node
whose value is a common scientific type opens a matplotlib plot.

Plot handlers (pdv/default_handlers.py)

• Default handlers for np.ndarray, pd.Series, pd.DataFrame, and xr.DataArray,
  registered during kernel bootstrap (register_defaults()), overridable by users.
• Each handler draws onto an explicit fig, ax (no plt.gca() ambiguity) and titles
  the axes with the tree path.
• Per-type behavior: np.ndarray 1D → line, 2D → imshow + colorbar, 0D/>2D → notice;
  pd.Series/pd.DataFrame → .plot(); xr.DataArray → .plot() (1D line / 2D
  pcolormesh / >2D histogram).
• Values that cannot actually be plotted (non-numeric Series/DataFrame,
  object-dtype arrays, …) close the half-built figure and print a [PDV] notice
  instead of raising — a raised exception would reach the renderer as an opaque
  internal.error. The final plt.show() is likewise guarded, so a
  backend-level failure degrades to a notice rather than escaping into
  dispatch_handler.

matplotlib & numpy dependencies

• matplotlib is now a declared core dependency of pdv-python (>=3.10.5).
  The floor is set by the dependency sweep: matplotlib 3.10.0 has a
  Path.__deepcopy__ infinite-recursion bug that breaks pandas.plot(), and
  3.10.5 is the first release with Python 3.14 wheels.
• numpy (>=2.2) is now also a core dependency. matplotlib already pulls
  numpy in transitively, and checksum.py / serialization.py import it
  directly — declaring it explicitly guarantees the >=2.2 floor the PDV code
  relies on instead of whatever matplotlib's transitive resolution would pick.
  The data optional extra is consequently now pandas-only; existing
  pdv-python[data] installs are unaffected since numpy still arrives via core.

xarray checksum support (pdv/checksum.py)

• tree_checksum() now content-hashes xr.DataArray and xr.Dataset nodes
  (KIND_DATAARRAY / KIND_DATASET) instead of falling back to repr(): dims,
  dtype, shape, raw values, coords, and attrs are all fed into the digest. Shared
  Dataset coords are hashed once rather than once per data_var.
• datetime64 / timedelta64 values (e.g. xarray time coordinates) take the fast
  buffer-protocol path alongside the plain numeric dtypes, rather than the slow
  repr(tolist()) fallback.

Test plan

• pytest pdv-python/tests/ — 482 passed, 2 skipped.
• pdv-python/scripts/sweep-deps.sh — all green (19/19 cells, exit 0:
  Python 3.10–3.14 × extras combos × highest/lowest-direct resolution; install +
  pytest pass in every cell). Re-run after the numpy core-dependency change — the
  lowest-direct cells now floor numpy at 2.2 as a direct dependency.
• New tests cover the graceful-notice path for non-numeric inputs, 2-D title placement,
  the xr.DataArray histogram fallback, and suppression of the overwrite warning on
  re-registration.