roofline: emit per-expert routing histogram (tokens_per_expert) for MoE launches

[roberteg16]

Summary

count_active_experts copies the MUL_MAT_ID ids tensor to host once per invocation to count the distinct routed experts. The same pass now also fills a per-expert token-count array. begin_op stores it per invocation and write_report emits it as tokens_per_expert (length n_experts, index e = routings to expert e, Σ = n_tokens*top_k) next to experts_used.

• No new device work or synchronization — the ids read-back already happens for experts_used; this replaces the seen[] bitmap with a counts[] array and adds one JSON field.
• Present only for MoE launches whose ids were read back; non-MoE rows are unchanged, and consumers that don't know the field ignore it.

Test plan

• ggml-cuda-roofline.cpp compiles clean with the project's exact build flags (host-only TU).
• End-to-end on gfx1151: GGML_ROOFLINE_OUT=... llama-bench -m granite-3.0-1b-a400m-Q4_K_M.gguf -p 64 -ngl 99 → all 71 MUL_MAT_ID rows carry tokens_per_expert, with len == n_experts, Σ == M*top_k, and nonzero == experts_used on every row.