Fix +/-inf in LSE returned by forward

[sgrigory]

Forward op was returning +inf in LSE for queries which have no keys to attend to, e.g. when K/V length happens to be 0. This diverges from the definition of LSE = log(exp(L1) + ... exp(L2)) which would give log(0) = -inf.
This PR fixes it, which allows feeding the output LSE directly into ops like merge_attentions without postprocessing.

pytest tests/test_flash_attn.py
...
======================================================================================== 268004 passed, 152064 skipped in 4404.00s (1:13:23) =========================================================================================

[tridao]

One issue I can see is that in the backward pass, if lse = +inf then exp(qk - lse) returns 0, which is what we want. If lse = -inf then exp would blow up.