tests: add MoE MMQ benchmark with routing-distribution generator

[roberteg16]

Summary

• Adds benchmark-moe-mmq, a standalone benchmark for the MMQ mul_mat_id kernel used by MoE.
• Generates controlled expert-routing distributions so the per-expert token load (which drives MMQ tile sizing) can be swept independently of everything else:
  • uniform - tokens spread evenly across all experts
  • single - all tokens concentrated on one expert (worst case)
  • zipf - power-law imbalance (--alpha)
  • concentration - sweep of tokens-per-expert aligned to 16 (active set shrinking from n_experts down to top_k)
• Reports tokens-per-expert stats and a roofline per row (arithmetic intensity, % of roofline ceiling, mem/comp bound); writes CSV.
• Defaults match Qwen3.6-35B-A3B (256 experts, top_k 8, Q4_K). Cycles distinct weight copies (--rotate, auto-sized from L2/MALL) so measurements reflect DRAM traffic rather than cache residency.
• Registered as a build-only target (llama_build), not a ctest, since it is a manual perf tool.

Test plan

• cmake --build build --target benchmark-moe-mmq
• ./bin/benchmark-moe-mmq --legend prints column/distribution help
• ./bin/benchmark-moe-mmq --dist all --proj both --m 512,2048 runs on the target backend and emits the table + CSV