mlxtop v3.1.2

Running requests now show their own generation speeds, llama.cpp reports a live
generation rate, and Linux detects AMD GPUs without extra tools.

Dashboards

• Each running request shows its own speed. While the runtime reports a
  decode speed per request, the Overview queue lists its running requests in
  its own rows, each with the tokens it has generated and its speed, such as
  8f31a2 1.7k · 41.0 tok/s, so a client's own tokens/s can be checked
  against it while the generation panel reports the whole server. oMLX
  reports these for every generating request; mlx-serve now does too, from
  each session's generated_tokens, read through the same stepping window as
  the server's decode speed; Splash has one while a request runs alone. A
  request without one shows — and keeps its row. The panel keeps its size,
  and the history returns when nothing is running.
• Prompt load shows the tokens each request generated. On a panel tall
  enough to keep four rows of bars, a row under the size labels shows each
  request's generated tokens. Whether the row is there depends on the panel's
  height alone, so the bars never shrink partway through a session.
• llama.cpp reports a live generation rate. Active decoding slots supply
  the current output rate, while slots still prefilling do not count toward it.
  The adapter accepts both object- and list-shaped /slots responses.
• Linux reads AMD GPU counters from amdgpu sysfs. Overview shows utilization,
  VRAM and temperature without a ROCm install. A runtime-suspended card stays
  listed without polling counters that could wake it on older kernels.
• mlx-serve session rates reset after a server restart. Reused request IDs
  cannot inherit a previous process's per-request speed.