fix: route group coordination requests to a single broker to prevent split brain

[klaudworks]

Summary

Closes #121.

When multiple consumers in the same group connect through different brokers, each broker independently tries to coordinate the group. This causes split brain: both brokers assign partitions, neither knows about the other's assignments, and messages get processed multiple times with no errors or warnings.

This PR fixes that by making brokers acquire an etcd lease before coordinating a group. Only the lease holder can coordinate — all other brokers reject the request with NOT_COORDINATOR, and the proxy retries on the correct broker.

What changed

Broker — Before handling any group operation (join, sync, heartbeat, leave, offset commit/fetch, describe), the broker tries to acquire an etcd lease for that group. If another broker already holds it, the request is rejected with NOT_COORDINATOR. If etcd is unreachable, the request is rejected with REQUEST_TIMED_OUT instead of silently proceeding (which was the old behavior that allowed two brokers to coordinate the same group simultaneously).

Proxy — The proxy watches etcd for group lease ownership and routes group requests to the owning broker. If the broker responds with NOT_COORDINATOR, the proxy invalidates its cache and retries on a different broker. The old detection logic used hard-coded byte offsets that broke on newer protocol versions and produced false positives on normal data. This is replaced with proper response parsing. Multi-group DescribeGroups requests are forwarded once without retry since different groups may live on different brokers — the client handles per-group errors natively.

Lease manager — The partition and group lease managers shared identical logic for session management, acquire/release, and etcd transactions. Both now delegate to a shared generic LeaseManager with thin type-safe wrappers on top.

Preexisting bugs fixed along the way

• Transient etcd errors during group lease checks were silently ignored, allowing two brokers to coordinate the same group. Now returns REQUEST_TIMED_OUT.
• Proxy failed to detect NOT_COORDINATOR on newer protocol versions (flexible headers shifted byte offsets). Replaced with proper response parsing.
• Proxy false-positived on normal response data that happened to contain the bytes for error code 16 (e.g., partition number 16, offset value 16). Structural parsing eliminates this.
• DescribeGroups with multiple groups caused a retry loop between brokers. Now forwarded once — the client handles per-group errors natively.
• Proxy retry loop aborted on the first connection failure instead of trying another broker. Connection errors now retry like other transient failures.

[novatechflow]

Nice one! Thanks :)