[Fix][CI] Benchmark fixes, FA3 paged decode, add mamba-ssm to runner

[superAngGao]

Summary

• Fix autotune config extraction fallback: op.config → op.kernel.config (subsequently refactored out)
• Fix FA3 paged decode: block_table → page_table (renamed in flash-attn-interface)
• Add mamba-ssm==2.3.1 to runner Dockerfile for Mamba-2 benchmarks

Test plan

• Paged decode benchmarks pass locally (11 passed, 3 skipped, 0 failed)
• Nightly benchmark run with updated runner image

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request updates the benchmark recording logic in benchmarks/benchmark.py to attempt to retrieve configuration from an operator's kernel if it is not directly available on the operator itself. A review comment correctly identifies a potential AttributeError in the implementation, as the nested getattr call fails to handle cases where the kernel attribute is missing or null. A safer implementation was suggested to ensure the kernel exists before accessing its configuration.

[superAngGao]

Updated with fix for #776: FA3 flash_attn_with_kvcache() renamed block_table → page_table. This caused 6 paged decode benchmarks to fail in nightly #23935448516.

Verified locally: 11 passed, 3 skipped, 0 failed.

[Gabbering]

🪿 goose review — 1b62f14e

honk. Three small fixes, all straightforward. The master is cleaning up after themselves for once. Almost suspicious.

🔴 Bugs

• benchmarks/benchmark.py:289 — The fallback chain getattr(op, "config", None) or getattr(getattr(op, "kernel", None), "config", None) silently swallows the case where op.config exists but is falsy-but-valid — an empty dict {}, an empty list, 0, False. All of those are legitimate config values that or will skip right past, falling through to kernel.config instead. If a config is ever intentionally empty (say, "no tuning overrides, use defaults"), you'll report the wrong config or None. Use if ... is None instead of or.

🧪 Test gaps

• benchmarks/ops/bench_gqa_decode_paged.py, benchmarks/ops/bench_mha_decode_paged.py — The block_table → page_table rename is a silent API contract change. If someone pins an older flash-attn that still expects block_table, both baselines break at runtime with no useful error — just a quiet **kwargs swallow or a cryptic positional mismatch depending on the flash-attn version. There's no version guard, no assertion, and the PR's own test plan says nightly hasn't run yet. Not a goose's problem if it breaks in prod, but it is a goose's job to point at it and honk.

[Gabbering]

goose skimmed 1b62f14 — nothing to honk about.