ROCm: validate the cooperative Q8_K quantizer on gfx1200 and gfx1201

[VikashLoomba]

Row: BACKEND-ROCM

Problem

Pull request #2472 closed #1876 after the cooperative Q8_K activation quantizer passed byte correctness, production-route mutations, same-binary A/B, and profiling on gfx1100. The implementation deliberately keeps the unset default on the legacy arm for gfx1200 and gfx1201 because neither architecture has equivalent runtime/default evidence.

The merged specification still needs an open owner for those external-hardware gates now that #1876 is closed.

Scope

Validate the merged Q8_K cooperative quantizer on actual gfx1200 and gfx1201 devices. For each architecture:

• compare all 292 BlockQ8_K bytes with the legacy GPU arm and CPU encoder;
• prove dense and grouped production reachability by mutation;
• run identical same-binary VT_ROCM_Q8K_BLOCK=0 versus =1 correctness and performance arms;
• capture runtime architecture resolution, toolchain, clocks, memory, contention, and profiler evidence;
• decide separately whether that architecture's unset default may select the cooperative arm.

Do not infer either result from gfx1100 source review, compilation, or measurement. Keep the current legacy unset default on each architecture until all of its gates pass.

Blocker

No currently authorized local device provides gfx1200 or gfx1201. This issue remains pending external hardware.

[VikashLoomba]

Issue #1876 was reopened after this duplicate was filed and already owns the exact gfx1200/gfx1201 Q8_K validation scope. Keeping one owner avoids splitting the same external-hardware gate across two issues.