metrics: spec-decode acceptance counters for the DFlash draft path

[xiaguan]

TL;DR: vllm:spec_decode_num_drafts/draft_tokens/accepted_tokens_total read 0; the DFlash speculative path (see #443) tracks draft/acceptance counts internally but nothing forwards them. Fill SchedulerStats.spec_decoding_stats from the qwen3 speculative executor. Part of #602.

Upstream records these as interval deltas (SpecDecodingStats in protocol/stats.rs), including per-position acceptance (num_accepted_tokens_per_pos) — the standard acceptance-rate dashboard panel.

Scope:

• Surface per-step draft/accept counts from the DFlash executor to the scheduler loop.
• Carry them on the published snapshot (delta-safe, same rule as the prefix-cache counters issue) and fill spec_decoding_stats in the bridge.
• Verify acceptance rate on /metrics matches the rate the DFlash perf test reports.

Blocked on nothing — the counters exist; this is plumbing. Coordinate with #443 so the metric lands with the roadmap line rather than after it.