Fix Metal model view overlap for q4

[Dango233]

Summary

Fixes q4 startup on machines where the model crosses Metal's per-buffer maxBufferLength, then documents q4 usage and benchmark results for a Mac Studio M2 Ultra with 192 GB RAM.

The original implementation used a fixed 672 MiB model-view overlap. The q4 GGUF has 1152 MiB expert tensors, so on this machine one tensor can straddle two mapped Metal views and fail startup with:

ds4: Metal model range 106.89..108.01 GiB is not covered by mapped model views

This PR fixes that by deriving the Metal model-view overlap from the largest tensor described by the GGUF metadata, with a CLI/server override kept as a diagnostic lower-bound escape hatch.

192 GB q4 setup

On the tested Mac Studio M2 Ultra with 192 GB RAM, q4 runs at --ctx 32768 after raising the iGPU wired-memory limit:

sudo sysctl iogpu.wired_limit_mb=188000

Benchmark

Single-run Metal CLI numbers using:

• --ctx 32768
• --nothink
• -sys ""
• --temp 0
• -n 256

Machine	Quant	Prompt	Prefill	Generation
Mac Studio M2 Ultra, 192 GB	q4	short	50.42 t/s	29.88 t/s
Mac Studio M2 Ultra, 192 GB	q4	3844 tokens	394.66 t/s	21.69 t/s

The long benchmark uses tests/test-vectors/prompts/long_code_audit.txt.

Test plan

• make ds4 ds4-server ds4_test
• ./ds4 --inspect -m ds4flash.gguf
• ./ds4_test --server
• q4 short benchmark command above
• q4 long benchmark command above