Add MOE variants and backend guide

[sunway513]

Summary

• Add comprehensive user-facing documentation for all MOE variants in AITER
• Cover Standard FusedMOE, Expert Parallel (EP), DP Shared Expert, Block-Scale FP8, MXFP4, End-to-End (E2E), and all quantization combinations
• Include routing options (TopK Softmax, TopK Sigmoid), activation support, and model-specific recommendations

Highlights

• Quick reference table helping users pick the right MOE config for their use case
• Quantization decision tree (BF16 → FP8 per-token → FP8 block-scale → MXFP4 → INT4)
• Full data type support matrix across ASM, CK, and Triton backends
• Pre-tuned model configs for Mixtral, DeepSeek-V2/V3, Qwen3-235B
• Practical API examples for each MOE variant (standard, EP, DP, block-scale, MXFP4)
• Performance tuning guide with environment variables and key parameters
• GPU architecture recommendations (MI300X vs MI350)

Test plan

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

🤖 Generated with Claude Code