Kernel workflow: Add FlyDSL gfx950 skills

[IzacharyI]

Summary

Add FlyDSL gfx950 expert skills for GEAK kernel_workflow.

This PR covers two kernel operator families:

1. Dense fp8 a8w8 blockscale GEMM
2. Grouped-MoE GEMM

These skills are consumed when use_expert_skills=true. They are advisory priors for kernel_workflow to reproduce and measure, not hard-coded replacements for correctness, parity, or A/B gates.

Modifications

Path	Change
perf_knowledge/expert_skills/skills/flydsl_fp8_blockscale_gemm/skill.md	Adds the gfx950 fp8 a8w8 blockscale GEMM kernel skill with validated q_up_proj, down_proj, and qkv_proj evidence. This is distinct from perf_knowledge/expert_skills/skills/flydsl_fp8_gemm_playbook/skill.md, which is a gfx942 e2e down-proj overlay playbook.
perf_knowledge/expert_skills/skills/flydsl_decode_moe_stage1_blkmap/skill.md	Adds the decode stage-1 grouped-MoE blkmap skill with phase/stage guardrails, descriptor timing scope, and runtime-signature checks.
perf_knowledge/expert_skills/skills/flydsl_prefill_moe_stage2_fp8partial/skill.md	Adds the prefill stage-2 grouped-MoE fp8-partial skill with phase/stage guardrails, scale-consistency checks, and relaxed-parity guidance.
perf_knowledge/expert_skills/index.yaml	Regenerates expert-skill selector metadata for the submitted validated skills.
kernel_workflow/roles/{tech_lead,engineer,deep_engineer}.md	Surfaces target backend/language authoring knowledge for cross-backend CK -> FlyDSL ports.
perf_knowledge/operators/scaled_quant_gemm/backends/{ck,flydsl}.md	Updates CK/FlyDSL scaled-quant GEMM guidance and expert-skill references for blockscale GEMM authoring.
perf_knowledge/operators/dense_gemm/{overview,numerics}.md	Updates dense GEMM overview and blockscale numerics references for CK -> FlyDSL workflow knowledge.
perf_knowledge/operators/dense_gemm/backends/flydsl.md	Updates dense GEMM FlyDSL backend guidance and expert-skill references.
perf_knowledge/operators/fused_moe_grouped_gemm/backends/flydsl.md	Updates FlyDSL grouped-MoE backend knowledge surfaced during workflow authoring.
perf_knowledge/quantization/block_scaling_mxfp.md	Updates block-scaling guidance to reference the current blockscale GEMM expert skill.

Skill Organization

The submitted skills are kept separate by workflow layer, GPU generation, and kernel scenario.

flydsl_fp8_blockscale_gemm is separate from the existing flydsl_fp8_gemm_playbook: the existing playbook is a gfx942 e2e down-proj overlay focused on capture-safe integration, while the new skill is a gfx950 kernel_workflow recipe for CK/production blockscale GEMM -> FlyDSL candidates.

The two Grouped-MoE skills are also kept separate: flydsl_decode_moe_stage1_blkmap targets decode / small-M stage-1 gate/up, while flydsl_prefill_moe_stage2_fp8partial targets prefill / large-M stage-2 down/reduce. They have different runtime signatures, validation scopes, and parity requirements.

Workflow Impact

• Expert skills are only consulted when use_expert_skills=true.
• Matched skills are advisory candidates for kernel workflow reproduction.
• Final acceptance still depends on on-box A/B measurement and parity validation.
• Default behavior with expert skills disabled is unchanged.
• CK -> FlyDSL ports now surface target backend/language knowledge earlier, so workflow agents can reuse validated GEMM recipes instead of rediscovering the porting rules.

Validation

Operator / Scenario	Skill	kernel_workflow / Director-verified result	Correctness
Dense fp8 blockscale GEMM / wide-N q_up_proj	flydsl_fp8_blockscale_gemm	~1.226x-1.237x	Pass
Dense fp8 blockscale GEMM / K-light down_proj	flydsl_fp8_blockscale_gemm	1.4541x	Pass
Dense fp8 blockscale GEMM / narrow-N qkv_proj	flydsl_fp8_blockscale_gemm	~1.05x	Pass
Grouped-MoE GEMM / decode stage-1 gate/up	flydsl_decode_moe_stage1_blkmap	1.0778x	Pass
Grouped-MoE GEMM / prefill stage-2 down/reduce	flydsl_prefill_moe_stage2_fp8partial	1.2019x	pass

Checks

• Static validation passes for all submitted skills.
• index.yaml matches skill frontmatter.
• git diff --check passes.
• No stale expert-skill links remain under perf_knowledge/expert_skills.
• No local validation artifact paths remain in the submitted skill files.
• No co-author trailers are present in the submitted commits.