actop 1.2.0

Added

• Per-process GPU attribution — the PWR column now covers GPU, not just CPU.
  A new gpu_registry.py module reads per-pid accumulatedGPUTime off each
  AGXDeviceUserClient via IOKit, sudoless. utils.get_top_processes() exposes
  a gpu_time_share alongside the existing cpu_time_share, and
  utils.attribute_power() combines both into the final watts value used by
  PWR and SORT_POWER. Completes Tier 2 of the feature-gap roadmap. Documented
  in DESIGN-system.md §5.7.
• scripts/ane_load.py — a CoreML-based Apple Neural Engine load generator for
  verifying that the ANE gauge reports power/percent correctly (the ANE reads
  0% (0.0W) when idle because it is power-gated). Builds an fp16 conv stack in
  memory, pins compute units to CPU+ANE, and loops inference.
• New ane optional-dependencies extra (coremltools, numpy) for the load
  generator. Kept out of the dev extra so Linux CI (pip install -e ".[dev]")
  stays lean and unaffected.
• README: DVFS residency comparison row, a Troubleshooting FAQ entry explaining
  the expected idle ANE 0% reading, and a Development note documenting the
  ane extra + scripts/ane_load.py.

Fixed

• Guarded native ctypes library loads in gpu_registry.py, smc.py, and
  ioreport.py under sys.platform == "darwin" (matching native_sys.py's
  existing pattern) — an unguarded load in gpu_registry.py was crashing
  import actop on non-Darwin, breaking CI's ubuntu-latest matrix.