cuda : add opt-in MoE cache telemetry env and early-router diagnostics

[ddvnguyen]

What

Small, opt-in debug tooling for the MoE expert cache and the experimental
GGML_CUDA_MOE_EARLY_ROUTER path. No behavior change when the env vars are unset.

1. GGML_CUDA_MOE_MM_DEBUG=1 now enables the existing mm-debug telemetry.
   Previously g_moe_cache_mm_debug could only be set through the C API
   ggml_backend_cuda_moe_set_debug_mm(), which only the tests call, so
   moe-cache: / moe-grouped-decode: / moe-cache-tensor-* output was
   unreachable from llama-server.
2. configure_early_router() one-time dormant diagnostics. When the early
   router is enabled but does not reach DECODE_GROUPED, it now prints the
   reason once (e.g. outcome-not-decode-grouped outcome=0, or
   dormant awaiting captured graph).
3. Server stats dump fix. The print_timings() stats call was behind
   #ifdef GGML_USE_CUDA, which is not defined in dynamic-backend builds
   (the CUDA backend is a shared object), so ggml_backend_cuda_moe_log_and_reset_stats()
   was never invoked from the server. It now uses the backend proc-address,
   matching llama_model_free().

Why

Needed to answer "why doesn't GGML_CUDA_MOE_EARLY_ROUTER engage?". With these
changes the telemetry shows the answer directly.

Usage

GGML_CUDA_MOE_EARLY_ROUTER=1 GGML_CUDA_MOE_MM_DEBUG=1 \
llama-server -m MODEL --moe-expert-cache-size N ...
# grep the server log for: moe-early-router:, moe-grouped-decode:, moe-cache:

Findings from the investigation (for reference, not fixed here)

• configure_early_router() requires execution->outcome() == DECODE_GROUPED
  (moe-cache.cu:8780).
• On plain decode the main-context execution certificate leaves
  flags = NONE (llama-context.cpp:4155-4178 only stamps REQUIRED_GROUPED
  for DRAFT/MTP, or the MAIN execution_intent branch used by speculative
  verification), so the plan is classified PREFILL_LEGACY and the early
  router can never engage.
• With MTP (grouping required) the candidate groups fail capability:
  required_unsupported=95 at --moe-expert-cache-size 2, still 63 at
  --moe-expert-cache-size 32; covered=0 and calls=0, so no grouped plan
  is built.
• Side effect observed: the legacy cache scales with size — apex + MTP went
  from 10.15 tok/s (slots=2) to 16.68 tok/s (slots=32).

Base: feat/763-reconcile-qwen4exp-mtp (the early-router code is not present on
baseline-flash-next), so this PR is only the debug tooling.

Assisted-by: opencode