Summary

Partially addresses #545. The sampler unit tests (test/unit_tests/samplers/test_samplers.py) downloaded the real BNCI2014_001 MOABB dataset (subject 4) in their windows_ds and target_windows_ds fixtures. These tests only depend on the dataset structure — the number of recordings, the number of windows, and the subject/session/run/target metadata columns — not on real EEG signals, so the download is unnecessary.

This replaces the two fixtures' data source with small synthetic data built from the existing braindecode.util.create_mne_dummy_raw helper (the util pointed to in the issue). The windowing calls and all test logic are left unchanged.

Why this is safe

The synthetic data reproduces the structure the samplers rely on:

• 3 recordings with distinct subject/session/run, so RecordingSampler groups them into the expected number of recordings;
• recordings long enough (52 s) that the fixed-length windows span the tau_pos/tau_neg range used by RelativePositioningSampler;
• 48 trials per recording across 2 alternating classes, so BalancedSequenceSampler has enough consecutive windows and at least two target classes (this covers the n_windows=40 parametrization).

Result

• All 19 tests in the file pass (12 non-distributed + 7 mp.spawn/gloo distributed).
• The data-loading cost is gone, and the tests no longer require network access or moabb. For example, test_recording_sampler went from ~11.5 s (real download) to ~3.0 s (synthetic; the remainder is just import overhead).

Scope: only test/unit_tests/samplers/test_samplers.py. The other files listed in #545 can follow in separate PRs.