Add MLA variants and backend guide

[sunway513]

Summary

• Add comprehensive user-facing documentation for all MLA (Multi-head Latent Attention) variants in AITER
• Cover Standard Decode, Persistent Decode, Standard Prefill, Persistent Prefill, Sparse MLA (Top-K), and Fused Operations (BMM+RoPE+Cache)
• Include backend support matrices (ASM vs Triton), data type coverage, GQA ratio support, KV cache layouts, and RoPE handling

Highlights

• Quick reference table helping users pick the right MLA variant for their use case
• Decision tree for backend selection (prefill vs decode, persistent vs standard, sparse vs dense)
• Data type matrices per variant and backend (BF16, FP8, FP4/MXFP4)
• GQA ratio support table including ASM persistent mode's simulated ratios (32-112)
• KV cache layout guide covering standard and 3-buffer FP8 layouts
• Practical API examples for decode, persistent decode, prefill, sparse MLA, and fused cache operations
• GPU architecture support summary (MI300X vs MI350 vs portable Triton)
• Performance tuning guide with split-K auto-tuning details and key parameters

Test plan

• Review report accuracy against current source code
• Verify all referenced API functions, source files, and kernel configurations exist

🤖 Generated with Claude Code