perf(flydsl): select validated Kimi-K3 KDA schedule

[JohnQinAMD]

Summary

Select the validated gfx950 Kimi-K3 KDA group64 default schedule:

• rows_per_wave: 1 -> 2;
• cu_count: 240 -> 256;
• keep weight_cache_modifier=2.

The kernel, quantization format, ABI, and explicit override controls are
unchanged. A regression test records the selected wrapper defaults.

This is stacked on perf/kimi-k3-kda-group64-gfx950.

Correctness and performance

The existing group64 campaign passed all changed-input checks bitwise. Two
independent schedule runs saved 0.035622 and 0.032652 ms/token
respectively; their mean is 0.034137 ms/token.

This schedule is intentionally not assigned standalone endpoint credit. It is
one member of the indivisible Wave40 bundle, whose three exact
TP8/B1/8K+1K/FP8-KV/no-spec trials improved the immutable Wave38 parent from
12.194182 to 11.886058 ms/token (82.006322 to
84.132180 tok/s/GPU). GSM8K first-100 was 100/100.

Validation

• Ruff: passed;
• Python compile: passed;
• focused wrapper/schedule tests in the exact Wave40 image: 10 passed;
• no NVIDIA source or dispatch path changes.

[JohnQinAMD]

Superseded by #11. The validated rows-per-wave 2 / CU256 / cache-modifier 2 schedule is now frozen directly in the standalone group64 PR, together with the extended contract tests and current endpoint evidence.