fix: gMLP uses full bias instead of truncated bias

[Mr-Neutr0n]

Bug

The gMLP SpatialGatingUnit.forward method correctly slices the projection bias to match the current sequence length (bias[:n]) for the causal case, but then passes self.proj.bias (the full, unsliced bias) to F.linear instead of the local bias variable. This causes a dimension mismatch when the sequence length is shorter than neox_args.seq_length.

Fix

Changed F.linear(gate.transpose(2, 1), weight, self.proj.bias) to F.linear(gate.transpose(2, 1), weight, bias) so the correctly sliced bias is used.