[Perf][Kimi K3] Support internal prefill checkpoints in Triton KDA

[gcanlin]

Follow-up to #52789 and #53614: add 64-token-aligned internal prefill checkpoints to the Kimi-K3 Triton KDA backend to avoid an extra full-model prefill pass.

Test results

16×B200 (2 nodes), TP8/DP2/EP16/DCP8, Triton KDA prefill, FlashInfer MLA, prefix caching enabled. 64 requests per workload, 1 output token, seed 42; cache reset before every workload. Latest completed paired round shown: round 2 except 128K/c16 (round 1). Off disables Triton's checkpoint capability (num_prefill_checkpoint_blocks=0, prefill_checkpoint_alignment=None) on the same source.

Input	Concurrency	Off mean TTFT (ms)	On mean TTFT (ms)	Reduction	Off P95 (ms)	On P95 (ms)
Random 8K	16	3593.4	3339.9	7.1%	4191.7	4213.7
Random 32K	16	10121.3	8999.8	11.1%	12524.3	12221.3
32K prefix + 2K suffix	1	1662.5	1343.2	19.2%	2179.3	1812.3
32K prefix + 2K suffix	16	8362.0	7141.1	14.6%	12874.5	11253.9
128K prefix + 2K suffix	1	3656.0	2790.1	23.7%	6051.7	5584.3
128K prefix + 2K suffix	16	28484.1	19156.9	32.7%	46044.2	35347.5

1,472/1,472 requests in completed workloads succeeded. Single-token outputs differed in 6/704 A/B pairs; repeat-to-repeat differences were 3/320 (off) and 4/384 (on).

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[gcanlin]

@ZJY0516 @yewentao256 Because FlashKDA has the accuracy problem, we're using triton KDA temporarily so migrate the optimization for TTFT from FlashKDA.