Add MacBook Pro M5 Max (128 GB) Metal throughput report

[lixiangnlp]

Summary

Adds a five-run ds4-bench performance report for MacBook Pro Apple M5 Max,
128 GB (Metal, IQ2 weight-mostly), using the exact command from the README:

./ds4-bench -m ds4flash.gguf \
  --prompt-file bench/promessi_sposi.txt \
  --ctx-start 2048 --ctx-max 65536 \
  --step-incr 2048 --gen-tokens 128

Five back-to-back sweeps with a passive cool-down between runs to keep one
hot sweep from dominating the headline numbers. The PR ships:

• bench/m5_max_128gb_report.md — host/model setup, per-run summary, and a
  per-frontier table with mean ± σ across the five runs.
• bench/results/run{1..5}.csv + bench/results/run{1..5}.log — raw outputs
  so any aggregate in the report can be regenerated.

Headline numbers (n = 5, Metal)

ctx	prefill mean (σ)	gen mean (σ)
2048	369.7 t/s (1.7)	31.46 t/s (0.07)
16384	261.9 t/s (6.8)	28.50 t/s (0.80)
32768	222.4 t/s (13.5)	24.29 t/s (1.27)
65536	161.5 t/s (5.9)	20.30 t/s (1.13)

Decode falls ~35% from 2k to 64k (31.5 → 20 t/s); prefill plateaus at ~160 t/s
past ~50k. Mid-context rows show the highest run-to-run variance — that window
of every sweep lands in the thermally-loaded phase, and run 4 specifically was
the hottest pass.

Test plan

• ./ds4-bench invoked with the README long-context command, 5 times
• Raw CSV/log captured for each run and committed under bench/results/
• Aggregates in bench/m5_max_128gb_report.md regenerated from the
  committed CSVs