Use `/proc/PID/mem` to access sandbox memory

[ValentaTomas]

• [ ] Test with both versions of used FC
• [ ] Add explicit handling for NoopMemory or implement more general memory mapping lookup

---

[!NOTE] Create memfile diffs by reading sandbox memory directly from /proc/PID/mem via a new mapped memory view, with UFFD providing the mapping.

• Memory diffing:
  • Switch MemoryDiffCreator to read from a generic io.ReaderAt and use a new process memory view instead of a tmp memfile.
  • In Sandbox.Pause, attempt snapshot with /dev/null (fallback to memfile if unsupported), obtain FC PID and UFFD mapping, build memory.View, and pass it to diff writer.
• UFFD integration:
  • Extend MemoryBackend with Mapping(ctx) and implement in Uffd (exposes mapping) and NoopMemory (returns nil).
  • Refactor userfaultfd.Userfaultfd to use m *memory.Mapping (rename ma), add Mapping() accessor, and update offset lookups.
• Memory mapping/view module (internal/sandbox/uffd/memory):
  • Add Mapping with GetOffset and GetHostVirtAddr, Region helpers, and View (implements io.ReaderAt over /proc/PID/mem), including specific errors.
  • Add comprehensive tests for mapping (offset/host-virt), view reading across regions/gaps, and UFFD-driven reads/writes/fault behavior.
• Misc:
  • Make block.Tracker.Has unexported (has) and update tests.
  • Update test utils to return detailed diff errors for byte slices.

^{Written by Cursor Bugbot for commit 7d148a050ea0c618166fb70b86448f9f1a93ab48. This will update automatically on new commits. Configure here.}

[ValentaTomas]

Waiting for https://github.com/e2b-dev/infra/pull/1385 merge.

[ValentaTomas]

@codex review

[ValentaTomas]

cursor review

[ValentaTomas]

Current blocker is the NoopMemory mapping—the actual maps we should be able to get from /proc/PID/maps, but if we want to improve just the non-build pause we can let that one be handled normally.

[ValentaTomas]

It is failing on integration test, because they are not using the custom FC.