AI spam

[Xavrir]

When several options share the same parameter name (a "feature-switch"
group, e.g. a plain --custom plus a flag_value --fetch that maps to the
same function argument), invoking only one of them produced the wrong value.

The parser stores command-line values keyed by the shared name. Because
options the user did not give are processed last, the un-invoked option read
the sibling's parsed value out of the parser results and overwrote the value
the invoked option's callback had already produced — so --fetch alone
yielded the raw sentinel instead of the callback's result.

This change tracks the parameters the parser actually saw
(Context._invoked_params) and, in Parameter.consume_value, skips adopting a
command-line value when this option was not invoked but a sibling sharing its
name was. The un-invoked option falls back to its own envvar/default.

This also fixes the secondary report: the "callback on both options" workaround
no longer calls the callback twice.

fixes #2786

Checklist (CONTRIBUTING.rst)

• Added a regression test that fails without the change
  (tests/test_options.py::test_flag_value_group_callback_not_overridden).
• Added a CHANGES.rst entry under the unreleased 8.5.0 section.
• Code-level comments added referencing the issue. (No public API/docstring
  change, so no .. versionchanged:: needed.)
• Full test suite passes locally (1669 passed, 24 skipped, 1 xfailed).

AI usage disclosure

This change was prepared with the assistance of an AI coding agent; the diagnosis,
patch, and regression test were reviewed and verified locally by the author.