fix(builder): add gRPC keepalive, stream timeout, and reconnection

[angerman]

Summary

The gRPC bidirectional stream between hydra-builder and the queue runner silently dies under certain conditions. When this happens:

• The builder process keeps running but receives no new work
• The queue runner shows the builder's job slots as occupied indefinitely (sinceLastPing growing to 50,000+ seconds)
• The only fix is a manual builder restart

Root cause

1. No HTTP/2 keep-alive on the tonic channel — dead TCP connections go undetected
2. No timeout on stream.next().await — the builder blocks forever waiting for a message that will never arrive
3. No reconnection — on stream error the builder exits instead of reconnecting

Changes

1. HTTP/2 keep-alive (10s interval, 20s timeout, active while idle) on all Channel::builder paths — detects dead TCP connections at the transport layer
2. 90-second receive timeout on the stream read loop — the queue runner sends pings every 30s, so 90s without any message reliably indicates a dead connection
3. Reconnection loop in main.rs — transient stream errors now trigger a 5s backoff + reconnect instead of process exit. Only API version mismatch is fatal.

Testing

Deployed to 10 darwin builders connected to a queue runner via IPv6. Previously these builders would go silent for 10-14 hours before manual intervention. With this patch, dead connections should be detected within ~90s and the builder should auto-reconnect within ~95s.

Fixes #1674