fix(client): enforce MaxConcurrentClusterConnects on physical dials

[SunSi12138]

Closes #661

Summary

• define MaxConcurrentClusterConnects as a coordinator-wide cap on concurrent physical child IClientTransportFactory.ConnectAsync attempts
• bind one shared dial limiter to initial child slots and runtime Add/Replace candidates
• route fixed, static-cluster, and dynamic-cluster dial paths through the shared boundary, covering initial connect, reconnect, pool expansion, and dynamic endpoint generations
• release permits exactly once on success, failure, or cancellation without touching the ordinary RPC hot path
• preserve fix(client): decouple runtime cluster add from readiness #660 running Add semantics: Start -> publish -> background readiness convergence; Add does not wait for remote Ready

Tests

Adds deterministic multi-cluster coverage for:

• initial child supervisors respecting the physical dial bound
• running Add publishing while its dial waits for a permit
• reconnect sharing the permit with a runtime-added child
• Replace candidates sharing the same physical dial boundary
• Stop cancelling permit owners/waiters without shutdown deadlock

The branch is based on the current dev head (5fac7a0ef6c96b2aa092160c2c89559695bf9b15).