fix(trtllm_mha): clamp page_table to k_cache page range to prevent SWA crash

[pyc96]

Summary

Defensive clamp that prevents the deterministic CUDA Warp Illegal
Address crash in fmhaSm100fKernel_*SlidingOrChunkedCausal* when
running Gemma-4 + --attention-backend trtllm_mha + MTP +
summarization workloads.

Stacked on #3 (debug trap). Staged on pyc96/sglang only.

What this fixes

Without this PR, the same workload that the bounds trap (#3)
catches in 30 s on E4B crashes the SGLang server with
cudaErrorIllegalAddress and SIGQUIT. With this PR, the same
workload completes cleanly.

How

In trtllm_mha_backend.py::forward_decode and forward_extend, right
after _get_layer_page_table and before the flashinfer kernel call,
clamp every page index to [0, k_cache.shape[0] - 1].

page_table = self._get_layer_page_table(layer, forward_batch)
num_pages_in_cache = k_cache.shape[0]
if num_pages_in_cache > 0:
    page_table = page_table.clamp(min=0, max=num_pages_in_cache - 1)

Three lines per call site, two call sites.

Benchmark — Gemma-4-E4B-IT, trtllm_mha, MTP, summarization 8 k/1 k × 80

metric	before this PR	after this PR
outcome	CRASH ~30 s in	COMPLETED
peak output token throughput	n/a	4,032 tok/s
median TTFT	n/a	1,437 ms
median TPOT	n/a	12.16 ms
accept length	n/a	1.61

Benchmark — Gemma-4-26B-A4B-IT, trtllm_mha, MTP, summarization 8 k/1 k × 80

metric	triton + Patch 1+2 (best safe baseline)	trtllm_mha + this PR	Δ
output throughput	1,097 tok/s	1,832 tok/s	+67 %
median TPOT	37.87 ms	24.97 ms	−34 %
median TTFT	8,763 ms	2,887 ms	−67 %
accept length	2.76	1.69	−39 % (see limitation below)

This unlocks the trtllm_mha attention backend for Gemma-4 MTP, which is
otherwise unusable.

Quality — MMLU @ 500 questions (Gemma-4-26B-A4B-IT, seed 0, temp 0)

server	accuracy
Patch 2 baseline (triton + MTP)	0.706
trtllm_mha + this PR	0.718
vLLM nightly (for comparison)	0.710

Within MMLU sampling noise; no regression.

Known limitation

Accept length drops from 2.76 (triton) → 1.69 (trtllm_mha + this PR)
on 26B summarization. Investigation:

• The clamp replaces OOB page indices with the LAST valid SWA page.
• For positions in the sliding-window range, the kernel will use that
  page's K/V instead of the correct one, producing slightly wrong
  attention values, which lowers MTP draft acceptance.
• This is a defensive safety net, not a complete fix. The
  underlying off-by-one in either full_to_swa_index_mapping or the
  SWA paged allocator's edge cases needs upstream investigation
  (filed as Patch E in humanize/source-idea-ledger.md).

For workloads where the lower acceptance is acceptable (the ~50 %
throughput improvement still significantly beats the triton baseline),
this fix is a net win.

Cost

One clamp() per kernel call. Few microseconds per forward. No
measurable performance impact.

Tests

No new unit test — the test is the reproducer:

agent-pad/runs/.../crash_repro/repro_e4b_bounds.sh         # used to crash
agent-pad/runs/.../crash_repro/repro_e4b_trtllm_eager_fix.sh   # used to crash, also fixed

Both reproduce no-crash behavior with this PR applied.

---

CI States

Latest PR Test (Base): ❌ Missing run-ci label -- add it to run CI tests.
Latest PR Test (Extra): ❌ Blocked -- run-ci is required first.