fix(docker-sandboxes): preserve runner keepalive lifetime

[joecklau]

Summary

• Keep the Docker Sandboxes managed keepalive alive after the provisioning-attempt context is canceled.
• Still terminate and reap the keepalive when startup itself is canceled.
• Add regression coverage for both post-start and during-start cancellation.

Root cause

The pool cancels a per-attempt provisioning context immediately after provisionOne returns. Docker Sandboxes used that context with exec.CommandContext for its long-lived sbx exec ... sleep infinity keepalive, so successful provisioning canceled the keepalive and the sandbox eventually stopped underneath long jobs.

Validation

• Full Go test suite passed.
• Focused and race tests for Docker Sandboxes keepalive behavior passed.
• Windows cross-compilation passed.
• Live HKAGE E2E attempt 7 ran for 19m33s, including a 16m15s Playwright step, without self-hosted runner communication loss. The workload concluded with three Playwright failures after 211 passing tests, and all artifact, cleanup, and post steps completed.