Fix unsafe client selection in ClientPool

[damilolaedwards]

GetClient's round-robin branch read the shared cursor and indexed the current candidate slice with it before checking that the cursor was still in range for that slice. Because the candidate set can be smaller on a later call than it was when the cursor last advanced (a different client group, a type exclusion, a client toggling enabled), this could index past the end of the slice and panic. The cursor is now clamped before it is read, and the duplicate round-robin logic that lived under both the explicit case and the default fallback is merged into one path.

Also fixes the client health check writing goodClients without holding selectionMutex, while GetClient only reads it under that lock. A write with no lock at all gives the reader no real protection, so the two could race on the slice header and hand back a stale length paired with a reallocated backing array.

Adds tests covering round-robin selection across a smaller group, a shrinking pool, the byIndex control path, and a concurrent write/read check for the lock fix.