feat(cli): time CDC transform stage separately from send

[aatran14]

why

The transform/embed stage is the dominant cost of interest in the CDC pipeline i.e. the embedding round trip. But currently it has no means of being recorded.

So far the system does an excellent job of keeping the measurements simple and concise like:

• whole-batch duration (which is just total time)
• turbopuffer latency
• replication lag

This PR entertains the integration of transform cost. Thus, we could measure transform / total time to measure how much the transform is soaking the CDC pipeline.

add

• observability.rs: cdc_transform_duration: Histogram<f64> on Metrics, built as
  puffgres.cdc.transform_duration_ms in build_metrics
• streaming.rs: wrap transform_batch in process_config_events with
  Instant::now() and record elapsed ms into that histogram

Live CDC only. Backfill is not covered yet (core::run_backfill takes no metrics param). Observability only, so no behavior change.

testing stuff

• cargo check -p puffgres-cli and cargo test -p puffgres-cli
• With OTLP enabled, confirm transform_duration_ms appears alongside the other CDC metrics
• the dumbest sanity test you can do on your machine is verify the the signal measures = ~0 under --provider none. The reason why this check make sense is when a real embedding provider is wired up, the transform makes the network call to the embedding model, which takes O(us)-O(ms) depending on regional overhead. Decoding the row is microseconds (us). The turbopuffer write is also a network call and is already measured via turbopuffer_latency