speed-bench: add M5 Max (Q2+Q4) results

[thomasjohnflaherty]

Benchmark results for Apple M5 Max (128 GB unified memory), Metal backend, mixed Q2+Q4 DeepSeek V4 Flash GGUF, at commit efdadd4.

Canonical sweep from the speed-bench README: ctx 2048..65536, step 2048, 128 greedy decode tokens per frontier. Headlines: 485 t/s prefill / 39 t/s gen @2k, 238 t/s prefill / 21.5 t/s gen @65k.

CSV is in the current ds4-bench output format (includes the gen_first_ms / steady-state columns); plot_speed.py renders it as-is.

[thomasjohnflaherty]

Heads-up that this overlaps with #584, which also adds M5 Max results at the same two paths (speed-bench/m5_max.csv and m5_max_ts.svg), so the two would collide on identical filenames.

They're not redundant, though: #584 is pure IQ2XXS full-residency, this PR is a mixed Q2+Q4 GGUF. The kvcache_bytes column is identical across both, so KV precision and the ctx 2048→65536 sweep are apples-to-apples — the divergence is purely the weight quant. The two curves cross: this run has higher prefill throughout (485 vs 340 t/s @2k), while #584's decode pulls ahead past ~32k. Having both quant points on the same machine class seems useful.

I've renamed mine to m5_max_q2q4.csv / m5_max_q2q4_ts.svg so both merge without collision; #584 can keep m5_max.*. Happy to fold into a single combined file with a quant column instead if you'd prefer.

Methodology note for comparability: these are steady-state numbers (the gen_steady_tps column) with a warm first token (gen_first_ms ~28 ms @2k), all greedy, --gen-tokens 128, at efdadd4 (current upstream main).