ESMFold2: fold(lm_dropout=0) does not turn the LM dropout off #418

The fold() docstring says lm_dropout defaults to 0.3 and that 0/None disables it, but 0 leaves the checkpoint's own rate in place: _lm_dropout_context returns early for any falsy value. The released biohub/ESMFold2 and biohub/ESMFold2-Fast configs ship lm_encoder.lm_dropout = 0.25 with per_loop_lm_dropout = True, so a fold asked for no dropout still drops 25 % of the LM pair channel in every loop:

from esm.models.esmfold2.processor import _lm_dropout_context

with _lm_dropout_context(model, 0):
    model.config.lm_encoder.lm_dropout  # 0.25, not 0

The helper's own docstring says "None/0 is a no-op", so the two docstrings disagree about what 0 should mean. For anyone using lm_dropout=0 as a control, the fold is not the dropout-free one it reads as.

What I would expect: 0 disables the dropout for the call and the config is restored afterwards, None keeps the checkpoint's setting, and a positive value overrides it, on both the release and the experimental architecture. If you would rather keep 0 as a no-op, the fold() docstring needs the fix instead.

I have the first version with tests on both architectures on a branch (diff) and can open it as a PR if that is the behaviour you want.