feat(downstream): surface cache_read/cache_creation on streaming responses (closes #52)

[arniesaha]

Summary

Closes #52.

#49/#50 added Anthropic cache_control injection and surfaced cached_tokens on the non-streaming response path. The streaming path (streamAnthropicToOpenAI) was still dropping the cache buckets entirely — AgentWeave and clients saw cached_tokens = 0 even when hit rate was ~99%. This PR closes that observability gap.

• streamAnthropicToOpenAI now captures cache_read_input_tokens and cache_creation_input_tokens from message_start.message.usage, and defensively from message_delta.usage (last value wins, matching the existing output_tokens handling)
• Trailing usage chunk emitted on success in OpenAI include_usage shape: { choices: [], usage: { prompt_tokens, completion_tokens, total_tokens, prompt_tokens_details? } }
• prompt_tokens rolls cache_read + cache_creation in, matching toOpenAIResponse — billable prompt size is now symmetric across streaming and non-streaming
• prompt_tokens_details.cached_tokens emitted only when Anthropic actually reports cache fields — plain requests stay clean
• StreamResult widened with cacheReadTokens / cacheCreationTokens; Anthropic adapter sets two new OTel span attrs (prov.llm.cache_read_input_tokens, prov.llm.cache_creation_input_tokens) on both streaming and non-streaming paths, only when >0

Notes for reviewers

• Chunk ordering follows the existing emitDownstreamResponseAsSse convention: finish_reason chunk → usage chunk → [DONE]. The issue's pseudo-code had them flipped; I went with the codebase-consistent order since any OpenAI client written against include_usage tolerates both.
• Span attrs use conditional spread (...(x > 0 ? {...} : {})) so they're absent rather than set-to-zero when cache is disabled or inactive — matches the pattern already used for callerAgentId.
• No new config flag. MUX_ANTHROPIC_PROMPT_CACHE=false naturally makes the new code emit a plain usage chunk with no prompt_tokens_details, since Anthropic won't return cache fields when we don't send cache_control.

Test plan

• npm test — 92/92 passing (5 new tests covering cache-on/off usage chunk, last-value-wins on message_delta, cacheReadTokens/cacheCreationTokens in StreamResult, log-event fields); 2 existing chunk-count assertions updated for the new trailing usage chunk
• npm run check — clean
• Deploy to prod, confirm cached_tokens > 0 on the trailing SSE chunk of agent-max turn 2+
• AgentWeave: confirm prov.llm.cache_read_input_tokens span attribute appears on streamed anthropic spans

🤖 Generated with Claude Code