Race: get_or_create_client can create duplicate clients for same context

[ndenev]

Summary

get_or_create_client has a check-then-create race for the same context.

Affected code

• src/kubernetes/client.rs (get_or_create_client)

Root cause

Current flow:

1. read lock check for existing client
2. release lock
3. create client
4. acquire write lock and insert

Parallel callers for the same context can both miss step 1 and both create clients; last insert wins.

Impact

• Primarily inefficiency/redundant work (duplicate kube client creation/handshake).
• Can add startup latency under parallel discovery.
• Lower correctness risk than discovery overwrite race, but still undesirable.

Proposed fix

1. Add a second check under write lock before insert (double-checked locking pattern).
2. Optionally use per-context once/mutex to guarantee single creation path.

Tests to add

1. Concurrent calls to get_or_create_client for same context should create at most one underlying client.
2. Verify all callers receive a valid client and no errors.