mlxtop v3.0.0

mlxtop now covers local LLMs beyond Apple Silicon: NVIDIA GPUs and Windows,
plus the mlx-serve and Splash runtimes.

• Windows (x86_64) support: memory, commit charge, processes and runtime
  detection through the Windows APIs, NVIDIA GPUs through nvidia-smi, logs in
  %LOCALAPPDATA%\mlxtop, a PowerShell installer (scripts/install.ps1) and a
  release zip. Windows exposes no paging counters through those APIs, so paging
  rates show as unavailable rather than zero.
• NVIDIA findings: the assessment names a model partly or fully on CPU
  (Ollama placement), a card with full VRAM (97%, held until below 94%), and
  thermal, power-brake or hardware clock slowdowns on a busy card. The GPU
  panel's STATE column shows the clock-limit reason, and full VRAM turns the
  card's memory reading yellow.
• Add Splash (MLXTOP_PROVIDER=splash, default port 8000), detected by its
  Metal engine process. It polls /status for the model, each running request's
  prompt and output counts with prefill progress, waiting requests, a generation
  rate from the engine's decode counter, the prefill speed of its prefill
  commands, prefix-cache reuse, Metal memory against its limit, DFlash draft
  acceptance, server TTFT p50, and released weights, thermal state and memory
  pressure. An API key goes in MLXTOP_PROVIDER_API_KEY.
• Add mlx-serve (MLXTOP_PROVIDER=mlx-serve, default port 11234) with process
  detection. It polls /metrics for queues, a live generation rate that counts
  tokens of requests still running, a prefill rate that follows the prompt being
  processed now, interval and cumulative cache reuse, and MLX active memory. It reads
  /metrics.json for the model and the running requests' prompt, cached and
  output counts; sessions that only hold a prefix cache are not listed. The
  server needs --metrics, and an API key goes in MLXTOP_PROVIDER_API_KEY.