fix: zero left-padded tokens before the GDN recurrence

[shiaho777]

GatedDeltaNet.forward received attention_mask and did not use it. A causal conv and the delta-rule update both carry state forward, so left-padding tokens change the real tokens that follow them. The loss mask removes those positions from the loss and does not clear the state.

Padding keys are the positions masked for every query, the same reduction gpt_model uses (attention_mask.all(dim=(1, 2))). Those hidden states are multiplied by zero before the projection. A mask whose shape is not [batch, seq] is left unchanged. Packed sequences keep using cu_seqlens and are not masked here.

Checked with python3 -m pytest tests/test_gdn_padding_keep.py. A [1, 1, 3, 3] mask with the first key padded produces keep [0, 1, 1]. A mask of the wrong length returns None. flake8 is clean.