Add SWITCH kernels write-up: performance vs vLLM's Punica LoRA path

[lastras]

Summary

• Adds docs/SWITCH_KERNELS.md, documenting the SWITCH fused-LoRA kernel backend (one fused base+shrink GEMM plus a Triton in-place expand kernel) against vLLM's three-launch Punica LoRA path.
• Covers balanced-rank and heterogeneous-rank decode speedups (1.2-1.4x and up to 5.9x), absolute throughput, the idle-adapter dispatch gap (and why it's a dispatch-policy difference rather than a kernel one), prefill, 8B results, and methodology.
• Adds the five supporting figures under docs/figures/.

Test plan

• Ran the repo's link validator (.pre-commit/validate_links.py) against the new doc — clean.
• Cross-checked every quantitative claim against the source benchmark tables.
• Verified the CLI example (compose_granite_switch.py --adapters/--output, vllm.entrypoints.openai.api_server) still matches current flags on main.