Harden Python sandbox transient retries

[mstolarzblaxelai]

Summary:

• Adds shared transient-reset retry handling for idempotent sandbox read/list and upload PUT paths.
• Keeps non-idempotent process execution and cleanup operations single-attempt.
• Adds local loopback transport-fault coverage for real httpx connection drops.

---

Note

Adds a shared transient_retry module with exponential-backoff retry logic for transport-level failures (connection resets, GOAWAY, timeouts). Wraps all idempotent sandbox read/list operations and upload PUT paths in both the async and sync clients. Non-idempotent operations (process start/stop/kill) are intentionally left single-attempt. Adds two new settings (BL_FS_PART_RETRIES, BL_SANDBOX_READ_RETRIES) to configure retry budgets per-operation class.

^{Written by Mendral for commit af560da.}

[mendral-app]

LGTM

The retry logic is correct — is_transient_reset_error properly gates retries on transport-level errors only, the _has_http_response_status guard prevents retrying real HTTP error responses, and closures correctly re-create BytesIO objects on each retry so multipart uploads don't replay an exhausted stream. CI failures (JSONDecodeError on HTML 403 in create_sandbox.py and missing sandbox-openai model) are environment-level issues in code untouched by this PR — safe to re-run.

_{Tag @mendral-app with feedback or questions. View session}