fix(runtime): retry transient sandbox startup failures

[bearsyankees]

No description provided.

[greptile-apps]

Greptile Summary

This PR adds retries for transient sandbox startup failures. The main changes are:

• Configurable startup attempts with exponential backoff.
• Transient transport error detection across exception chains.
• Cleanup of failed sessions before each retry.
• Immediate retry cancellation when cleanup fails.
• Tests for retries, cleanup failures, and staged directory lifetime.

Confidence Score: 5/5

This looks safe to merge.

• A failed teardown now stops the retry loop before another sandbox is created.
• The new test confirms that the original startup error is raised and only one session is created when teardown fails.
• No blocking issues were found in the updated code.

Important Files Changed

Filename	Overview
strix/runtime/backends.py	Adds sandbox startup retries and stops creating sessions when cleanup of a failed attempt does not complete.
tests/test_backend_retries.py	Covers transient retries, non-transient failures, failed cleanup, repeated attempts, and staged directory cleanup.

_{Reviews (2): Last reviewed commit: "fix(runtime): fail closed when sandbox t..." | Re-trigger Greptile}

[devin-ai-integration]

@greptile