Fix fence tracking for donated dynamic slice offsets

[michaelellis003]

Proposed changes

Fixes #3880.

Metal's compute_dynamic_offset can donate the dynamic-start input buffer to
its one-element offset output. The donated alias is currently registered as
an encoder temporary. During CommandEncoder::end_encoding(), temporary
cleanup removes that shared pointer from the input and output dependency sets
before producer fences are resolved, so a dynamic slice can use a stale start
after an encoder boundary.

Register the offset as a temporary only when its storage was newly allocated.
The donated path remains enabled, and the primitive input retains the shared
storage through command-buffer completion. This matches the CPU path and the
temporary-lifetime pattern fixed for vector SDPA in #3121.

Per review, the committed regression is now the public slice_update
reproducer from #3880 (no internal APIs).

Testing, on an M3 Pro (36 GB, macOS 26.2), base
8c28c38:

• Committed regression: on pristine main it fails 10/10 fresh processes
  with default buffer limits (the 512 MB reduction crosses the per-buffer
  memory limit, so the command-buffer boundary lands between the start's
  producer and the slice) and 10/10 with MLX_MAX_OPS_PER_BUFFER=0; with
  the fix, 0/20 in both modes.
• Public slice reproduction from [BUG] Metal slice_update can use a stale array-valued start across command buffers #3880, fresh processes with
  MLX_MAX_OPS_PER_BUFFER=0: pristine 20/20 wrong; fixed 0/50 wrong.
• Same reproductions under METAL_DEVICE_WRAPPER_TYPE=1
METAL_DEBUG_ERROR_MODE=0: pristine 20/20 wrong; fixed 0/20 wrong.
• Full C++ suite on Metal: 264 test cases, all passing, plain and under the
  Metal validation wrapper.
• Full Python suite: 810 tests, OK (46 skipped).
• pre-commit run --all-files: passing.

Checklist

• I have read the
  CONTRIBUTING
  document
• I have run pre-commit run --all-files to format my code / installed
  pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature
  works
• I have updated the necessary documentation (if needed — no public API
  change)

[zcbenz]

Adding a test relying on internal platform-dependent API is quite tricky, I think using your test in the original issue should be enough.

[zcbenz]

Thanks!