metal: pread pool dispatch stats and IO-tier pinning for streaming experts

[jasontitus]

Two independent pieces from the streaming-IO investigation in #533,
refiled without the withdrawn default change (see that thread for the
cross-machine data).

1. metal: report pread pool dispatch stats in the timing summary —
   one extra line under the existing
   DS4_METAL_STREAMING_EXPERT_TIMING_SUMMARY gate: dispatches, tasks,
   bytes, average dispatch wall time, effective queue depth (sum of
   per-task pread ms / pool wall ms) and delivered bandwidth. Sample from
   a ctx-2048 streaming prefill (Flash IQ2XXS imatrix, 33 GiB cache
   budget, --simulate-used-memory 64):

   ds4:   streaming pread pool dispatches=2031 tasks=27279 tasks_avg=13.4 workers_avg=8.5 bytes=59.94 GiB wall_avg=1.286 ms qd_avg=5.82 pool_gbps=24.64 task_gbps=4.23
   

   This line is what surfaced both halves of the metal: stop expert-miss readahead racing the pread pool (+13% GLM streaming decode) #533 result (the
   glm5.2-line eval-thread stall and the main-side prefetch win); it
   makes streaming IO regressions visible without dtrace.

2. metal: keep expert pread threads on the important IO tier — reader
   threads inherit the parent QoS, so under taskpolicy -b or background
   launchd jobs every expert miss pays the throttled IO tier. Pin them to
   user-initiated QoS + IOPOL_IMPORTANT;
   DS4_METAL_DISABLE_STREAMING_EXPERT_PREAD_QOS opts out. Foreground
   runs are unaffected (measured neutral on this machine and
   independently on an M5 Pro in the metal: stop expert-miss readahead racing the pread pool (+13% GLM streaming decode) #533 thread).

Machine: M1 Ultra Mac Studio 128 GB / 8 TB SSD, macOS 15, Metal backend
(one machine only).

Tests: make clean && make (zero warnings); ds4_test --server and
ds4_test --metal-kernels OK; --logprob-vectors pre-existing ERR
(long_code_audit) with the IQ2XXS imatrix quant, identical on unpatched
main.