Replace linear channel-name lookups with dictionaries

[bruAristimunha]

Reference issue (if any)

None. Found while profiling a read → pick → filter → epoch → average workflow.

What does this implement/fix?

Three places in mne/_fiff matched channel names by scanning a list, which is
quadratic in the channel count:

• _make_projector() called col_names.index(name) inside a loop over
  ch_names. A set was already used for the membership test, but the lookup
  next to it still scanned.
• pick_channels() did three list scans per included channel:
  name in ch_names, name not in exclude, and ch_names.index(name).
• _picks_str_to_idx() called info["ch_names"].index(pick) once per pick.

Each becomes a single dict lookup. ch_names is already asserted unique in
pick_channels, and duplicate col_names already raise in _make_projector.
In _picks_str_to_idx the map is built with setdefault, so it keeps
list.index()'s first-match semantics exactly, even though ambiguous names are
rejected further down anyway.

Medians of 3 interleaved A/B rounds against a pristine main worktree:

	128 ch	306 ch	1000 ch
pick_channels	2.1x	3.6x	10.2x
_picks_to_idx	1.4x	2.3x	6.4x
_make_projector	1.6x	2.9x	3.3x

End to end, apply_proj() on a 392-channel raw goes 21.7 → 17.8 ms. The
speedup grows with channel count, so it matters most for whole-head MEG and
high-density EEG.

Correctness

• 24/24 projector matrices bit-identical to main across six real MEG files,
  with and without bad channels.
• 700 fuzzed pick_channels / _picks_to_idx cases identical, including
  duplicate names, absent names, ordered=True/False, and the exceptions
  raised.
• pytest mne/_fiff mne/io mne/tests mne/preprocessing mne/channels:
  3506 passed, 0 failed.

Two tests are added. They pin name-based mapping specifically: the projector
test uses col_names permuted relative to ch_names, so a positional match
annihilates the wrong direction and fails. Mutation-tested — substituting
position for name in either function is caught, as is dropping the bads
filter, ignoring exclude, and silently skipping missing names.

Additional information

AI disclosure: I directed the work and reviewed and tested everything; Claude
Code (Claude Opus 5) did the profiling that found these, made the edits, and
ran the A/B measurements and fuzzing under my direction.

[larsoner]

Nice, I think it actually makes the code cleaner in a couple of places, too!