feat: session affinity + prompt cache for multi-turn conversations

[icebear0828]

Summary

• Session affinity: route all turns of a conversation to the same account via responseId → entryId mapping, fixing previous_response_id breaking across account rotation
• prompt_cache_key: send per-conversation UUID to enable backend prompt caching — Turn 4+ sees 93% cache hit rate (was 0% before)
• Missing fields: forward service_tier, include on WebSocket path (were silently dropped)
• Request-level monitoring: affinity hit/miss, payload size, usage stats

Verified

Turn     in    out  cached  cache%
Turn 1    87   3698       0    0.0%
Turn 2  3772    157       0    0.0%
Turn 3  3948    146       0    0.0%
Turn 4  4111   1519    3840   93.4%  ← prompt cache hit

Changes

File	Change
session-affinity.ts	New: responseId → (entryId, conversationId) map with 4h TTL
proxy-handler.ts	Affinity lookup/record, generate conversationId, set prompt_cache_key + include
response-processor.ts	Pass through onResponseId callback (was discarded)
account-lifecycle.ts	acquire() supports preferredEntryId hint
codex-types.ts	Add prompt_cache_key, include to request type
ws-transport.ts	Add service_tier, prompt_cache_key, include to WS message
codex-api.ts	Forward new fields on WS; stop stripping service_tier on HTTP

Test plan

• 1394 tests pass
• Session affinity: 10 unit tests (account mapping + conversationId tracking)
• Account acquisition: preferredEntryId passthrough
• E2E: 4-turn conversation → same account, affinity=hit, cached_tokens > 0