Add Mac Studio M2 Ultra q2 benchmark results

[Dango233]

Summary

Adds q2 benchmark results for a Mac Studio M2 Ultra with 192 GB RAM.

The new M2 Ultra rows are 5-run averages using:

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

The long benchmark uses tests/test-vectors/prompts/long_code_audit.txt, which reports 3844 prompt tokens and triggers chunked Metal prefill.

Results

Machine	Quant	Prompt	Prefill	Generation
Mac Studio M2 Ultra, 192 GB	q2	short, 5-run avg	51.86 t/s	30.72 t/s
Mac Studio M2 Ultra, 192 GB	q2	3844 tokens, 5-run avg	401.58 t/s	22.56 t/s

Benchmark Runs

Short prompt:

• Prefill avg: 51.86 t/s, sample sd: 0.66
• Generation avg: 30.72 t/s, sample sd: 0.04

Long prompt:

• Prefill avg: 401.58 t/s, sample sd: 0.62
• Generation avg: 22.56 t/s, sample sd: 0.07

[gmontana]

FWIW, M4 Max (40-core GPU), 128 GB numbers under the same protocol as this PR. The laptop chassis throttles ~30% over 5 back-to-back long runs, so I added 2 warmup passes and 60 s cooldowns between measured runs to stay off the throttling regime. No other deviation.

Machine	Quant	Prompt	Prefill	Generation
MacBook Pro M4 Max (40-core GPU), 128 GB	q2	short, 5-run avg	63.16 t/s	32.23 t/s
MacBook Pro M4 Max (40-core GPU), 128 GB	q2	3844 tokens, 5-run avg	302.33 t/s	24.84 t/s

Benchmark Runs

Short prompt:

• Prefill avg: 63.16 t/s, sample sd: 1.02
• Generation avg: 32.23 t/s, sample sd: 0.23

Long prompt:

• Prefill avg: 302.33 t/s, sample sd: 3.73
• Generation avg: 24.84 t/s, sample sd: 0.59

Happy to send a separate PR with the M4 Max row once the protocol is settled.