The previous layout nested the encoder and contextualizer under
`encoder_state_dict` / `contextualizer_state_dict` top-level keys, which braindecode's loader did not unwrap: `load_state_dict(..., strict=False)` silently matched 0 of 99 weights and every call to `BENDR.from_pretrained` returned a random model. The flat layout matches braindecode's `BENDR` key namespace (`encoder.encoder.*`, `contextualizer.*`, plus the random classification head), so `from_pretrained(..., strict=True, n_outputs=2)` now loads all 99 pretrained weights. See braindecode PR #992.
