v0.8.57: Normalize cached and reasoning token telemetry across all providers

[Hmbown]

Goal

Extend the provider usage-telemetry normalization started for the OpenAI Codex provider (#2955) to every supported provider, so the Harbor stream and benchmark harness report a consistent token schema: input, cached input, output, and reasoning tokens, with explicit null for metrics a provider does not report.

Current evidence

• v0.8.56: Build a Codex-parity token comparison harness #2952 notes the CodeWhale Harbor stream currently reports aggregate input/output only, while Codex CLI separately reports cached input and reasoning tokens — making parity comparisons lossy.
• v0.8.56: Align OpenAI Codex provider usage telemetry with Codex CLI #2955 fixes this for the OpenAI Codex provider only.
• DeepSeek's API reports prompt_cache_hit_tokens/prompt_cache_miss_tokens, and other providers (OpenRouter, Together AI Provider: add dedicated Together AI support #2906, etc.) expose their own cached/reasoning usage fields; none of these are currently normalized into one schema.
• 输入缓存命中率太低了 #1177 (low input cache hit rate) is hard to diagnose without per-provider cache-hit telemetry surfaced consistently.

Scope

• Define one normalized usage struct (input, cached_input, output, reasoning, plus provider-raw passthrough) and map each supported provider's usage payload onto it.
• Emit it uniformly in the Harbor stream and wherever per-turn usage is recorded, so the v0.8.56: Build a Codex-parity token comparison harness #2952 harness and future tooling read one schema regardless of provider.
• Document per-provider field mappings and which fields are unavailable (explicit null, never silent zero).
• Tests with recorded usage payload fixtures per provider.

Non-goals

• Changing what the TUI displays during long tasks (that UX is telemetry: agents need visible token context and resource usage during long tasks #2666).
• Improving the cache hit rate itself (输入缓存命中率太低了 #1177, v0.8.56: Reduce repeated transcript input in benchmark and exec turns #2956, PR refactor(prompts): decouple allow_shell from static system-prompt prefix #2949 cover that) — this is observability only.
• Adding new providers.

Acceptance criteria

• All currently supported providers map their usage payloads to the normalized schema, with fixtures proving the mapping.
• The v0.8.56: Build a Codex-parity token comparison harness #2952 comparison harness consumes cached/reasoning tokens for non-Codex providers without provider-specific parsing.
• A docs table lists each provider's usage-field mapping and gaps.

Related: #2955, #2952, #1177, #2956, #2666

Deferred to v0.8.57: depends on #2955 landing first and touches every provider adapter, so it should not block the v0.8.56 parity work.