Add Nx CPU benchmarks and refresh docs

[goodhamgupta]

Summary

• add Nx CPU backend baselines (Torchx/EXLA) to benchmark generation, CSV schema, and plots; add a warmup to avoid first-run skew
• update the benchmark runner to include Nx CPU backends when available
• refresh the README with benchmark environment details and updated latency/speedup ranges, and note the CPU baseline
• runtime safety tweaks: gate AVX2 SIMD to avoid illegal instructions, reduce variable-length doc copying, remove deprecated rustler init args, and make Nx a required dependency

Testing

• mix test
• OPENBLAS_NUM_THREADS=1 RAYON_NUM_THREADS=8 mix run bench/generate_plots.exs
• uv run bench/plot_benchmarks.py