LaBraM: keep the pretrained time embedding at every window length, restore mean pooling - #1155

#1153 asked whether LaBraM's temporal embeddings are handled correctly. In decoder mode they are: _PatchEmbed averages the grouped channels, so there is one token per temporal patch (Copilot's clarification and test, kept here). The actual defect was in the default tokenizer mode, together with a readout default that had drifted from the original.

What was wrong

• Time embedding sized to the window. temporal_embedding had one slot per patch plus one, while the released weights hold the original 16 absolute slots (time_embed in modeling_finetune.py). They only fit 15-patch windows. On master, Labram.from_pretrained("braindecode/labram-pretrained", n_times=800) fails with size mismatch for temporal_embedding: [1, 16, 200] vs [1, 5, 200], so any other window length has to drop the pretrained slots and learn them from scratch.
• Readout default. [EHN] LaBraM automatic channel reordering and multiple fixes #931 made use_mean_pooling=False the default so that the pretraining checkpoint loads strictly. That readout is the [CLS] output, which LaBraM's pretraining loss never uses; the docstring says True, and the original fine-tuning script sets use_mean_pooling=True (fc_norm of the mean patch token).

What changed

• In tokenizer mode, temporal_embedding keeps the original 16 absolute slots (patch p still uses slot p), with extra slots only for windows longer than 16 patches. A load hook copies every slot a checkpoint holds, so the released weights load at every window length, and checkpoints saved with the previous one-slot-per-patch layout load with identical outputs. When a window has more patches than the checkpoint has slots, a warning names the slots that keep their initialization. Decoder mode is unchanged.
• use_mean_pooling=True is the default again. A pretraining checkpoint (with norm, without fc_norm or head, like the released weights) loads into the mean-pooling model as in the original fine-tuning script: norm is unused and fc_norm keeps its initialization. A checkpoint fine-tuned with the [CLS] readout fails to load into the default model instead of being converted silently; pass use_mean_pooling=False to keep that readout.
• The decoder temporal-embedding test now passes use_mean_pooling=False explicitly. It replaces the final norm with Identity to read the tokens, which only holds for the [CLS] readout.
• whats_new: the time-embedding fix under Bug fixes, the readout default under API and behavior changes.

Copilot's decoder clarification (variable names, _PatchEmbed docstring, test) is unchanged.

Evidence

• Identical to the original code. braindecode vs modeling_finetune.py from 935963004/LaBraM, both with the released weights; the braindecode side loads the Hub checkpoint with a plain strict load_state_dict. Max |diff| is 0 for every token, for the [CLS] readout and for the mean-pooling readout, at 18 geometries: 22-channel (BCI IV 2a), 19-channel (10-20) and 64-channel montages × 1, 3, 4, 5, 10 and 15 patches. Labram.from_pretrained(..., n_times=800) gives mean pooling, 16 slots and max |diff| 0.
• Tests. 6 new tests (9 cases). 6 cases fail on the previous head of this PR (9224f3e) and pass here; the other 3 guard compatibility (15-patch windows, old one-slot-per-patch checkpoints, rejecting [CLS]-fine-tuned checkpoints). pytest test/unit_tests/models -k labram: 81 passed, 9 skipped. Full test/unit_tests/models: 2753 passed, 184 skipped, 0 failed (same code as this head; only a docstring and whats_new were reworded afterwards).
• No new ruff findings (ruff 0.14.9, as in .pre-commit-config.yaml).

Fixes #1153