docs(speed-bench): add generated benchmark summary

[dutifulbob]

Opened on behalf of Onur Solmaz (osolmaz).

Summary

Speed benchmark CSVs are easy to add but hard to compare at a glance.
This change adds a generated summary table to speed-bench/README.md.
It also adds a small Python updater so the table can be rebuilt whenever CSV files change.

What Changed

The speed bench folder now has one command that rebuilds the summary from the checked-in CSV files.
The generated README block has explicit begin and end markers so future updates replace only that section.

• Added speed-bench/update_summary.py using only the Python standard library.
• Appended a generated benchmark summary table to speed-bench/README.md.
• The table includes best generation, generation at 32k context, average generation, best prefill, prefill at 32k context, and average prefill.
• CSV files with UTF-8 BOMs are handled with utf-8-sig.

Testing

I tested the updater directly and checked that the generated Markdown is clean.
This is a documentation/script-only change; I did not run model inference benchmarks.

• python3 speed-bench/update_summary.py
• PYTHONDONTWRITEBYTECODE=1 python3 -m py_compile speed-bench/update_summary.py
• git diff --check -- speed-bench/README.md speed-bench/update_summary.py

Machine/backend/model quant: not applicable for this docs-only change.
Notable failures: none.

Risks

The risk is low because this does not touch inference code or benchmark CSV data.
The main limitation is that the summary only reflects CSV files in speed-bench/, not one-off numbers in the top-level README.

[osolmaz]

Feel free to refactor/reimplement

I thought it would be cool to have a single place to see all the perf results