Fix Reformer LM head bias not applied since the weight-loading refactor (#50811)
Fix Reformer LM head bias dropped since the weight-loading refactor
Since #41580, `ReformerOnlyLMHead.decoder` is created with `bias=False` and the line that set
`self.decoder.bias = self.bias` was removed, so the trained output bias stored in checkpoints as
`lm_head.decoder.bias` is reported as an unexpected key and never applied. Use `bias=True` as BERT
does since #41580, and re-record the 2 seeded random-init LSH tests whose forward outputs change
with this change.
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>