[EventHubs] Harden CBS token refresh retry and recovery

## Summary

The proactive CBS token refresh added by #7308 keeps long-lived Event Hubs clients authorized, but its recovery behavior still has several gaps. Refresh attempts use one fixed connection-wide cadence, refresh failures lose their structure, one slow audience can delay another, and an authorization failure can require a second user operation before a rebuilt connection succeeds.

This issue tracks the follow-up work that makes CBS refresh and foreground recovery bounded, failure-aware, and deterministic. Rust AMQP parity is outside this workstream.

## Scope

- [ ] #7376
- [ ] #7378
- [ ] #7379
- [ ] #7377

The implementation remains internal. Any public configuration or API change requires a separate API review.

Issue #7330 covers the related low-level ambiguity where `PutTokenForAudience` raises `AuthenticationException` for every non-`Ok` CBS result. Issue #7379 preserves the classified result through the refresh scheduler without adding Rust parity to this workstream.

## Completion criteria

- [ ] Every linked sub-issue is complete.
- [ ] A transient CBS refresh failure can recover before the current authorization expires.
- [ ] A stale authorization can rebuild once within the same send or receive operation.
- [ ] A permanent credential failure remains bounded and preserves its actionable error.
- [ ] Retry and recovery waits honor `Azure::Core::Context` cancellation.
- [ ] Deterministic offline tests cover retry success, retry exhaustion, cancellation, and multiple audiences.

This work changes authentication behavior and requires explicit security-focused review.

Start by reading linked sub-issues #7376, #7378, #7379, and #7377, then inspect the proactive CBS refresh work from #7308 and the Azure::Core::Context cancellation path. Done means bounded, cancellation-aware retry and recovery with deterministic offline tests for success, exhaustion, cancellation, and multiple audiences, while preserving actionable permanent credential errors.

Written by the indexing model from the issue text.