Support LMContext with temperature=0

[gabegrand]

There are cases (e.g., testing / debugging) where it's useful to sample with temperature=0.

However, currently, LMContext requires 0 < temperature < infinity:

hfppl/hfppl/distributions/lmcontext.py

Line 112 in a191fca

temp (float): temeprature for next-token distribution (0 < temp < float('inf'))

Setting it to 0 results in a zero-division error:

hfppl/hfppl/distributions/lmcontext.py

Lines 127 to 129 in a191fca

	self.next_token_logprobs = log_softmax(
	lm.next_token_logprobs_unbatched(self.tokens) / temp
	)

Is there a reason why in principle we can't support temperature=0?