Restore acceptance tests on supported Python versions - #1159

Co-authored-by: bruAristimunha <[email protected]>

On Windows, _parse_description_from_file_path normalises the mocked
POSIX file paths with os.path.normpath, so _get_header received a
backslash path and raised KeyError in
test_variable_length_trials_cropped_decoding (Windows 3.12/3.13 CI).
Normalise the lookup key to forward slashes inside the mock and add a
focused unit test that exercises both separators on every host.

test_trialwise_decoding is deterministic for a fixed seed, but it asserted
valid_accuracy[0] < valid_accuracy[-1] and valid_loss[0] > valid_loss[-1]
on a 30-trial validation split after 6 epochs. Those comparisons move by
one trial and failed for 40/160 seeds (25%), so a torch/BLAS/platform change
flips them; locally the fixed seed now gives 0.600 -> 0.567.

Keep the finite/range checks on all metrics and the improving train
loss/accuracy, and add:
- best train loss < 0.8 * first epoch (worst of 160 seeds: 0.651)
- final train_accuracy >= 0.65, i.e. 39/60 (worst of 160 seeds: 43/60)
- predict() accuracy on the train and valid splits equals the final
  history scores

The test now guards the training pipeline (labels, loss, optimizer,
predict and scoring), not generalization. It remains a network test and
is skipped in CI without --run-network.

The acceptance tests asserted first-vs-last history improvements that failed
for 23-32 of 50 seeds, and the cropped tests used nll_loss on logits (the
models no longer end in log_softmax), so their losses went negative and
their held-out accuracy stayed at chance.

- Use BNCI2014_001 subject 1 (left vs right hand), which the ubuntu 3.12
  CI job already downloads, instead of PhysioNet EEGBCI.
- Trialwise: 4-fold CV predicts all 288 trials once while held out; assert
  held-out accuracy >= 0.67, a binomial test against chance, and a
  shuffled-label control trained with the same pipeline <= 0.64. A second
  test fits one fold twice under deterministic algorithms and requires
  identical histories and probabilities.
- Cropped: train on session 1, predict the 144 trials of session 2 with
  cross-entropy and a cosine lr schedule; accuracy >= 0.75, control <= 0.66.
- EEG classifier and variable-length tests: cross-entropy, non-negative
  losses, predict/scoring consistency and measured train-loss checks.

Thresholds come from local sweeps that varied the init, batch-order and fold
seeds separately (trialwise real 0.708-0.812 over 231 runs, control
<= 0.604 over 275; cropped real 0.826-0.882 over 151, control <= 0.590
over 101).

The matrix jobs run pytest without --run-network, so the real-data
acceptance tests were always skipped. Add one ubuntu 3.12 job with
OMP_NUM_THREADS=1 that restores (without saving) the ubuntu 3.12 MNE data
cache and runs test/acceptance_tests with --run-network.

…rials

np.array() on ragged per-trial predictions raises ValueError on NumPy >= 1.24. Equal-length trials still return the same 3-D array.