fix: retry transport-sourced Cancelled errors and count them in metrics

[brucearctor]

Summary

When Code::Cancelled originates from a transport-level failure (tonic::transport::Error → hyper::Error, e.g. connection closed during an AZ outage), the SDK now treats it as Unavailable-class and retries using the standard bounded backoff budget instead of the previous single-retry goaway workaround.

Additionally, transport-level failures that never produce a grpc-status HTTP header are now counted in the request_failure metric with a status_code=TRANSPORT_ERROR label.

Fixes #1252

Problem

Three compounding issues cause silent activity completion drops during sustained network turbulence:

1. One-shot retry limit: The goaway-cancel workaround in retry.rs only allows a single retry. During sustained outages (not a single GOAWAY frame), that one retry burns immediately against a still-dead connection.

2. Silent drop: After the retry exhausts, activities.rs catches the forwarded error and logs warn!("Network error while completing activity") — no metric, no callback, no error propagation.

3. Metrics blind spot: GrpcMetricSvc::call() only increments request_failure when it can parse a grpc-status header from the HTTP response. Transport-killed calls return Err(...) at the HTTP level, so the counter is never incremented.

Changes

crates/client/src/retry.rs

• Replaced the one-shot have_retried_goaway_cancel flag with is_transport_cancelled() — a helper that inspects the error source chain for tonic::transport::Error → hyper::Error
• Transport-sourced Cancelled errors now flow through the standard backoff budget (same as Unavailable)
• True application/caller-initiated cancellations (no transport error in source chain) remain non-retryable per gRPC spec

crates/client/src/metrics.rs

• Added svc_request_failed_transport() method with status_code=TRANSPORT_ERROR label
• GrpcMetricSvc::call() now counts Err(...) responses (transport failures) in addition to Ok(...) responses with non-OK grpc-status headers

Testing

• All 11 existing retry unit tests pass
• Clippy clean with -D warnings

[Sushisource]

Just some minor comment cleanup. Thanks a lot for this, the idea is a good one.

[brucearctor]

committed/accepted suggestions.

Thanks @Sushisource !

[brucearctor]

I can't rerun, but seems like flaky CI test?