Prepare release v1.5.1 - #1023

Bump version from 1.5.0 to 1.5.1 and split the post-v1.5.0 changelog
entries (BandRotation, EMG2QwertyNet SpecAugment + return_features,
AmplitudeScale RNG crash fix, Jonathan Dan author link) into a new
'Current 1.5.1 (stable)' section, mirroring the pattern from the
v1.5.0 release commit.

PR #993 (1.5.0) removed the per-batch ch_names kwarg from
Labram.forward and tightened the __init__ chs_info check into a
ValueError. Downstream wrappers (neuroai's _LabramChannelWrapper)
were still calling Labram with non-canonical chs_info and a per-batch
ch_names subset, producing 7 test failures.

Restore the kwarg as keyword-only (via *); when provided it
case-insensitively maps each name to LABRAM_CHANNEL_ORDER and uses
those indices for the position-embedding bank. When None, fall back
to the 1.5.0 arange-over-canonical behavior. The __init__ canonical
check is downgraded to a UserWarning so wrappers can build the inner
Labram with their union channel set and resolve the subset per batch.

Update test_labram_rejects_non_canonical_chs to assert the warning
(renamed to test_labram_warns_on_non_canonical_chs). Document the
restoration under 1.5.1 'API and behavior changes' in whats_new.

Apply review feedback:
- Lift the ch_names -> canonical-index dict to module scope as
  _LABRAM_CANONICAL_INDEX instead of rebuilding it every forward
  call (~128 string ops + dict inserts saved per batch).
- Strip changelog narrative (the 'Restored in 1.5.1' / neuroai mentions)
  from the forward docstring and __init__ comment — that history lives
  in whats_new.rst, not in the runtime code.
- Tighten the non-canonical-chs_info warning: drop Sphinx :meth: /
  :class: roles (they don't render at runtime) and collapse to 4 lines.
- Drop the redundant 'x is already in canonical order' comment in the
  forward fallback branch — the arange already says it.
- Trim the length-mismatch ValueError to the equation alone.

Behavior unchanged; all 25 braindecode labram tests + 13 neuroai
test_labram tests still pass.

- forward(): move `*` after the return_* flags so only `ch_names` is
  keyword-only. `return_patch_tokens`, `return_all_tokens` and
  `return_features` stay positional, preserving back-compat for
  callers that used positional args before #993.
- forward(): when `ch_names is None`, validate `x.shape[1] ==
  len(LABRAM_CHANNEL_ORDER)` up front and raise a clear ValueError
  pointing to `ch_names=` or `InterpolatedLaBraM` instead of failing
  later with a confusing shape mismatch inside `forward_features`.
- forward() docstring: clarify that `ch_names` is keyword-only and
  only honored when `neural_tokenizer=True`; in decoder mode the
  position embedding is sequential.
- examples/plot_channel_interpolation: drop the redundant in-body
  `import warnings` (already imported at module top) and rewrite the
  "silently mis-align" wording — the model warns at construction and
  forward now raises ValueError on a non-canonical channel count.
- test_foundation_models: add focused tests for the ch_names path
  — subset forward, case-insensitive matching, unknown-channel
  ValueError, length mismatch, the new None+non-canonical guard,
  and back-compat for positional return_* flags.