Add exponential backoff and log suppression for scheduler disconnection

[phillipleblanc]

Summary

When the scheduler disconnects or becomes unavailable, the executor poll loop previously logged a WARN message every ~100ms, spamming logs during graceful shutdown scenarios (e.g., when shutting down the scheduler before the executor).

Before:

2025-12-15T23:02:44.486012Z  WARN ballista_executor::execution_loop: Executor poll work loop failed. If this continues to happen the Scheduler might be marked as dead. Error: status: Unavailable, message: "tcp connect error", details: [], metadata: MetadataMap { headers: {} }
2025-12-15T23:02:44.588781Z  WARN ballista_executor::execution_loop: Executor poll work loop failed...
[repeats every ~100ms]

After:

• First 5 failures: WARN with attempt count
• Subsequent failures: DEBUG level (reduces noise)
• Exponential backoff from 100ms to 30s between retries
• Info log when connection is restored

Changes

• Add exponential backoff (100ms initial, 30s max) when scheduler connection fails
• Reduce log level from WARN to DEBUG after 5 consecutive failures
• Log restoration message when connection is re-established
• Prevents log spam when scheduler shuts down before executor