CAS daemon no longer fsyncs analytics on every operation under build-time disk contention

The CAS daemon’s per-operation analytics database now opens with PRAGMA synchronous = OFF instead of NORMAL, so each storeMetadata insert is a page-cache append rather than an fsync. On hosted macOS runners under build-time disk contention this was stalling the cooperative pool and inflating CAS per-op latency to hundreds of ms (size-independent, like a flaky disk). Additionally, the daemon now records transfer_duration (network) and codec_duration (de/compression) separately from the total, so the remaining time can be attributed to daemon overhead.