feat(BUB-213): cross-provider USD cost normalization

[Bububuger]

Summary

• Adds packages/cli/src/runtime/shared/pricing.ts with an 8-model pricing table and computeCost(model, usage) that returns {input_usd, output_usd, total_usd} or undefined for unknown models
• usageAttributes in usage.ts gains optional model param; emits gen_ai.cost.* OTLP attributes when cost is computable
• turn.ts passes latestModel / assistant.model to usageAttributes at all 5 call sites
• thinking_tokens billed at output rate (Anthropic extended thinking)
• SPANORY_PRICING_<MODEL_SAFE> env var override for custom pricing
• 4 new fields registered in telemetry/field-spec.yaml
• 10 unit tests covering: known model pricing, unknown model → undefined, env override, version suffix stripping, zero tokens

exec-plan → docs/exec-plans/bub-213-usd-cost.md

Test plan

• npm run check — 0 errors (903 pre-existing warnings)
• npm test — 244 tests pass (40 test files), including 10 new pricing tests
• npm run telemetry:check — ok, 0 validation errors

[Bububuger]

审阅后，现不宜批准；有以下问题需先修：

1. packages/cli/src/runtime/shared/pricing.ts:23-25
   OpenAI 定价表有确定性错误。gpt-5.4 官方当前价格是 $2.50 / 1M input、$15.00 / 1M output，代码却写成了 10 / 40。这会把 GPT-5.4 成本放大约 4 倍。参考：https://openai.com/api/pricing/ 与 https://developers.openai.com/api/docs/models/gpt-5.4。

2. packages/cli/src/runtime/shared/pricing.ts:19-27
   仓内现有 Codex 数据实际使用的是 gpt-5.3-codex，而不是 gpt-5.4；但 pricing table 没有这个模型。可见 packages/cli/test/unit/normalize.spec.ts:24,38 与 packages/cli/test/fixtures/codex/sessions/session-a.jsonl:4,16。结果是当前 Codex 会话会直接 computeCost(...) === undefined，完全不产出 gen_ai.cost.* 字段，和此 PR 目标不符。OpenAI 当前 gpt-5.3-codex 价格为 $1.75 / 1M input、$14.00 / 1M output：https://developers.openai.com/api/docs/models/gpt-5.3-codex。

3. packages/cli/src/runtime/shared/pricing.ts:107-112
thinking_tokens 被无条件加到了 output_tokens 上。对 Anthropic extended thinking 这可能成立，但对 OpenAI/Codex 不成立：OpenAI 官方 usage 里 output_tokens_details.reasoning_tokens 是 output_tokens 的细项，且 max_output_tokens 明确“including visible output tokens and reasoning tokens”。现在这段会对 OpenAI/Codex 双重计费。参考：https://platform.openai.com/docs/guides/reasoning?api-mode=chat&lang=python。

补充：zero tokens 已覆盖，但 negative tokens / negative env override 仍可能产出负美元值；建议至少 clamp 到 0 或直接 skip，并补测试。