feat: add speed benchmark, interactive example, and RTX 4080 test

[MasahiroOgawa]

Objective

Add support for RTX-4080 GPU, and add visual test and interactive test.

Test result

• benchmark_text_generation_latency_visual.py successfully outputs this image;
• benchmark_speed_mamba123.py successfully outputs this image;

- examples/predict_next_token.py successfully outputs this response;

Summary

New files only — no modifications to existing code.

• benchmarks/benchmark_text_generation_latency_visual.py — Compares text generation latency across multiple Mamba model sizes, outputs visual chart
• benchmarks/benchmark_speed_mamba123.py — Compares Mamba1 vs Mamba2 vs Mamba3 forward/backward speed on identical workload, outputs visual chart
• examples/predict_next_token.py — Interactive next-token prediction with example outputs on launch showing what base LM prediction looks like (supports Mamba1 and Mamba2 pretrained models)
• tests/test_rtx4080.py — RTX 4080 (sm_89) test for pretrained inference and Mamba3 SISO module
• configs/rtx4080.json — Model presets sized for 12GB VRAM
• benchmarks/benchmark_README.md — Explains the purpose of each benchmark
• README.md — New sections for interactive example and speed benchmark

RTX 4080 findings

• Mamba3 SISO mode works (Triton kernels)
• Mamba3 MIMO backward requires >100KB shared memory per SM (H100/sm_90+ only)

Test plan

• Ran benchmark_text_generation_latency_visual.py on RTX 4080
• Ran benchmark_speed_mamba123.py on RTX 4080
• Ran predict_next_token.py with mamba-130m and mamba2-130m
• Ran test_rtx4080.py for pretrained inference and Mamba3 SISO fwd/bwd

🤖 Generated with Claude Code