fix(ep): AsyncLL slot assignment double-allocates when top-k does not divide warpSize

[TianDi101]

Problem

EpDispatchLowLatencyAsyncSendCopySlotAssign tiles a wavefront into
tokensPerWarp = warpSize / numExpertPerToken token groups. If top-k does not divide
warpSize, the trailing lanes form an incomplete group that resolves to the next warp's
first token:

warpSize=64, numEpt=6 -> tokensPerWarp=10
warp g   lanes 60..63 -> token g*10 + 10
warp g+1 lanes 0..3   -> token (g+1)*10     // same entry

Both warps take the atomicAdd branch, so one entry consumes two slots while
dispDestTokIdMap[i] keeps only one. The surplus slot is orphaned — SendCopyMultiBlock
never writes it, but RecvTransfer signals the inflated count, so the receiver reads
untouched memory as a token. Its index bytes give pe = id / numExpertPerRank out of
range, tripping the assert in RecvCopyMultiBlock.

Note this corrupts dispatch before it ever crashes: some tokens are routed twice with one
copy reading stale memory, so runs that complete without asserting are also affected.

Hidden until now because 64 % 8 == 0. DeepSeek-V4 uses top-k 6 (64 % 6 == 4), which
exposed it. Top-k 9 was already covered but _AsyncLLCombineOnlyTestCase skips
check_dispatch_result at non-multiple-of-8 top-k, so it could not catch this.

Fix

if ((inWarpTokIdx >= tokensPerWarp) || (tokenId >= args.curRankNumToken)) continue;

Active lanes only __shfl from baseLane + j < tokensPerWarp * numEpt, so masking these
lanes cannot disturb the dedup shuffle. No-op where top-k divides warpSize. This tiling
pattern appears at one site in the tree.

Test

test_dispatch_slot_assign_lane_tiling checks three top-k agnostic invariants, so it works
where the positional check cannot: received count matches the deduplicated host-side
expectation, source positions are unique, and each decodes to an in-range
(src_rank, src_id). Parametrized over top-k 8, 6, 9, 10, 12 (8 is the clean-tiling
control).

Verified on 8x MI355X:

New test, before fix	FAILED — Rank[0] received 587 dispatched tokens, expected 547
New test, after fix	5 passed
Full test_dispatch_combine_async_ll.py, after fix	194 passed, 233 skipped, 0 failed