fix(uffd): add retry with exponential backoff on source.Slice() error

[jakubno]

Transient Slice errors (network blips, temporary GCS/S3 failures) previously caused immediate sandbox termination. Retry up to 3 times with exponential backoff (50ms-500ms + jitter) before signaling uffd exit, giving transient errors a chance to recover.

[cursor]

PR Summary

Medium Risk
Changes the page-fault data path to add timed retries, which can affect sandbox startup latency and failure behavior under load. The retry loop also introduces timers and jitter per fault, so mis-tuned backoff could amplify resource usage if many faults hit transient errors simultaneously.

Overview
Adds a retry loop around source.Slice() when faulting in pages via UFFD: on transient read errors it now retries up to 3 times with exponential backoff (50ms–500ms) plus jitter, and only then signals failure/exit, logging attempt counts and returning an error that includes the total attempts.

^{Reviewed by Cursor Bugbot for commit 5fa2cb0. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

Two issues found in the retry loop in userfaultfd.go. Line 376: time.After creates a timer not cancelled when ctx.Done fires, leaking it until expiry. Use time.NewTimer and call timer.Stop on context cancellation. Line 369: rand.Int63n panics when the argument is 0 - safe with current 50ms base delay but fragile if sliceRetryBaseDelay is reduced below 2ns. Guard with: if half := int64(delay) / 2; half > 0.