[Cute,Sm100] allow for zero length sequences in hdim 256 kernels

[jayhshah]

We add support for zero-length Q and KV sequences for varlen mode in the sm100 hdim 256 kernels. Changes are as follows:

Forward: have softmax execute one dummy iteration to not hang for zero-length K. We also add row sum check for zero or NaN to not output NaN in this case.

Backward dQ: for guarding work tile per-iteration, check also that trip count is non-zero. If it is zero, write zero as the output.

Backward dKV: zero-length Q and K was nominally supported but the write zero logic was broken and yielded IMA; now fixed in the PR.

[Johnsonms]

PR fixes the bug.

Before patch (b11 baseline):

• forward OK → out.backward(g) → cudaErrorIllegalAddress in _flash_attn_bwd (sm100 hd=256 kernel)

After patch (jshah/hdim256-varlen-zero-lengths, commit 75db52f):

• forward OK
• backward OK
• grads finite, non-zero, and structurally correct — dk[0:2538] (the K rows paired with the zero-length Q segment) is
  exactly zero as expected, while dk[2538:] carries signal.