Optimize KDA prepare cu_seqlens scan

[lyppg]

In current varlen mode, every CTA of _flash_kda_fwd_prepare maps its global_tile_idx to a sequence index by linearly scanning cu_seqlens from the beginning. This PR precomputes a prefix sum of per-sequence tile counts once and have each CTA locate its sequence with an O(log N) binary search.
Result:
Tested on GB200, under varlen, N x 1024 tokens, H=96, D=128, bf16.

N	tokens	before (ms)	after (ms)	latency reduction	speedup
8	8,192	0.6884	0.6857	0.4%	1.00x
16	16,384	1.3377	1.2980	3.0%	1.03x
32	32,768	2.6996	2.4867	7.9%	1.09x
64	65,536	6.0072	4.8818	18.7%	1.23x
128	131,072	15.2356	9.8047	35.6%	1.55x
256	262,144	44.0999	19.8455	55.0%	2.22x