[AMD] Reuse KV gather indices across ASM context prefill layers

[zijiecode]

Motivation

The gfx950 ASM context-prefill path gathers paged K/V into contiguous buffers. It rebuilds the same gather indices at every full-attention layer, including four GPU synchronizations per call. In a TP2 trace, this takes about 35 ms of a 490 ms prefill chunk with 16k new tokens and 15 full-attention layers. This PR computes the indices once per batch and reuses them across layers.

Modifications

• Reuse the token indices and cumulative lengths already available in prefill metadata for ordinary extend batches.
• Cache the indices in ForwardMetadata. Speculative batches and unsupported metadata use the existing helper and bounds checks.
• Add SGLANG_GFX_ASM_CTX_GATHER_CHECK=1 to compare the reused indices with the original helper during validation.

Each layer still gathers its own K/V data. The attention kernel is unchanged.

Accuracy Tests

amd/Qwen3.5-397B-A17B-MXFP4-AttnFP8-V2, TP2/C8, 5-shot GSM8K, chat-completions, 16384-token output budget, temperature 0, real MTP.

Configuration	Strict-match	Flexible-extract
Baseline	97.88%	97.80%
This PR, chunked prefill 512 so that gsm8k prompts get chunked	97.65%	97.65%

Benchmarking and Profiling

amd/Qwen3.5-397B-A17B-MXFP4, MI355X, TP2/C32, HiCache, and MTP. Two 1200-second AgentX runs per configuration on the same GPU pair; the table shows arithmetic means of the per-run metrics.

Metric	Baseline average	This PR average	Change
Standard p90 interactivity (tok/s/user)	39.24	40.26	+2.60%
Throughput (tok/s/GPU)	547.10	547.50	+0.07%
TPOT mean (ms/token)	16.260	15.855	-2.49%
TPOT p50 (ms/token)	12.065	11.870	-1.62%
TPOT p90 (ms/token)	25.490	24.840	-2.55%

Checklist

• Format your code according to the Format code with pre-commit guide.
• Add unit tests.
• Update documentation.
• Provide accuracy and speed benchmark results.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ⏳ Run #35750753383
Latest PR Test (Extra): ❌ Run #35750753270
Latest PR Test (AMD ROCm 10): ⏳ Run #35750753475

[yichiche]

Only affect AMD path. LGTM.