[levanter] lm_eval uses unsupported `implementation` arg

[eric-czech]

lm-eval passes implementation to fused_cross_entropy_loss_and_logsumexp_penalty from levanter.models.loss here:

marin/lib/levanter/src/levanter/eval_harness.py

Line 284 in e78e1dc

implementation="xla",

and this isn't supported in any of the overloads of fused_cross_entropy_loss_and_logsumexp_penalty of levanter/models/loss.py#L160-L216.