AbstractCore v2.15.3

Ejecting an MLX model now frees all the memory it held, and the memory snapshot shows what MLX
still keeps.

Fixed

• Ejecting an MLX model now frees everything it held: the weights, its prompt and KV caches and the
  drafter, from every holder in the process, and MLX's allocator cache is cleared. The eject
  (abstractcore.providers.mlx_residency.eject_model) reports what it freed and any residual
  instead of claiming success. Who is affected: anyone who ejects models to reclaim memory —
  previously a gateway could hold tens of GB of Metal buffers with no model listed. On a 27B model
  the process footprint now drops from 22 GB to 0.65 GB after the eject; before, it stayed at 22 GB.
• A residency record for an MLX instance that released its hold while another instance in the
  process still keeps the weights now reads loaded, with provider_state: "resident_via_other_holders" and the held_bytes still in memory, instead of "not loaded".

Added

• The memory snapshot (get_memory_snapshot(), GET /acore/memory) reports MLX active, cache,
  peak and held bytes (device.mlx_*_bytes), the MLX models alive in the process and their holders
  (held), and, on macOS, the process's physical footprint (process.footprint_bytes), which
  includes Metal memory that RSS leaves out.