Add reproducible Mac Studio M3 Ultra benchmark submission

[nabbilkhan]

Summary

This PR adds a reproducible community benchmark submission package for Mac Studio (Apple M3 Ultra, 256 GB RAM), captured against commit 443194bca4491fae4400bae9dad2a0470692bdbf.

It is intended as a concrete contribution toward cross-chip coverage in #3.

What is included

• benchmarks/README.md with a lightweight submission format
• benchmarks/submissions/m3-ultra-mac-studio-2026-03-03/ containing:
  • README.md (environment + commands + key metrics)
  • metrics.json (machine-readable results)
  • commands.sh (exact repro commands)
  • raw/*.log + raw/system_info.txt + raw/upstream_commit.txt (auditable raw outputs)
• Root README.md pointer to community benchmark submissions and this M3 Ultra entry

Key results (20-step training runs)

Pipeline	Avg train	ANE TFLOPS	Total TFLOPS
train_large	81.2 ms/step	1.15	2.15
train_large_ane	71.4 ms/step	1.48	2.44
train_large_ane --no-ane-extras	123.8 ms/step	0.85	1.41
training_dynamic/train --scratch	115.4 ms/step	n/a	n/a

inmem_peak best observed: 8.08 TFLOPS (128x conv 512ch sp64).

Notes

• Serial/UUID identifiers are redacted in raw/system_info.txt.

• inmem_bench and sram_bench outputs are included as observed (FAIL(-1) for all rows on this clean setup), to keep the submission fully transparent.

• Disclosure: I am a contributor to Open Claw.