feat: retry transient 401/403 auth errors with exponential backoff

[Rolf Håvard Blindheim (rhblind)]

Summary

• Add configurable max_auth_retries option (default 3) so the producer retries transient 401/403 auth errors during Splunk rolling updates instead of immediately stopping the pipeline
• Retries use exponential backoff (base interval * 2^n) capped at 30s or the base interval, whichever is larger
• Emit [:off_broadway_splunk, :auth, :error] telemetry event on each auth error for observability
• Fully backwards compatible: existing producers get the new retry behavior by default; set max_auth_retries: 1 to restore the previous immediate-stop behavior

Closes #19

Test plan

• 4 validation tests for max_auth_retries option (negative, zero, default, positive)
• 4 unit tests for backoff_interval/2 (doubling, cap when base < 30s, cap when base > 30s, zero base)
• Transient 401 recovery test (2 failures then success)
• Exceeding max retries stops the producer
• Legacy behavior test (max_auth_retries: 1 stops immediately)
• Telemetry event emission test
• Existing 401/403 tests still pass unchanged
• Full suite: 45 tests, 0 failures
• Dialyzer, Credo, format checks pass

[Rolf Håvard Blindheim (rhblind)]

Approved