Add core operator documentation guides and update README

[sunway513]

Summary

• Add comprehensive documentation guides for 10 core AITER operators: Attention (MHA/PA), MLA, Fused MOE, GEMM, Quantization, Normalization, RoPE, KV-Cache, Elementwise/Activations, and Sampling
• Each guide covers available variants, backend support (ASM / CK / Triton), Python API examples, and performance tuning advice
• Update README with improved installation instructions (JIT dev mode, precompiled kernels, build variables) and a new Supported Operators table linking to each guide

Files

Guide	Operator
docs/attention_variants_guide.md	Multi-Head Attention, Paged Attention, Unified Attention, GQA/MQA
docs/mla_kernel_support_report.md	Multi-head Latent Attention — Triton vs ASM backends
docs/moe_variants_guide.md	Fused MOE — A8W8, A16W8, FP8 block-scale, MXFP4, 2-stage
docs/gemm_variants_guide.md	GEMM — A8W8, A16W16, A4W4, batched, DeepGEMM, Triton FFN
docs/quantization_guide.md	Quantization — FP8/MXFP4/INT4, per-tensor/token/block, SmoothQuant
docs/normalization_guide.md	RMSNorm, LayerNorm, GroupNorm — fused add/quant variants
docs/rope_guide.md	Rotary Position Embedding — SBHD/THD/2D/3D, NeoX & GPT-J
docs/kv_cache_guide.md	KV-Cache — paged/flash/MLA layouts, FP8/INT8, fused RoPE
docs/elementwise_activation_guide.md	Activations & elementwise — SiLU/GELU/SwiGLU, fused quant
docs/sampling_guide.md	Token sampling — greedy, random, top-k, top-p
README.md	Updated installation + Supported Operators table

Guide Structure

Each guide follows a consistent format:

1. Quick Reference table for fast lookup
2. Numbered sections covering variants, API, backends, data types
3. Decision tree for variant selection
4. Source files and test files tables for code navigation

[Copilot]

Pull request overview

This PR adds operator-focused documentation guides for core AITER components and refreshes the main README to improve installation guidance and provide a linked “Supported Operators” index.

Changes:

• Added multiple operator guides under docs/ covering API usage, backend support, and tuning notes.
• Added a consolidated “Supported Operators” table to the README linking to the new guides.
• Expanded README installation instructions (JIT dev mode, prebuild options, environment variables, Iris comms).

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
docs/attention_variants_guide.md	New guide documenting attention variants and backend selection.
docs/mla_kernel_support_report.md	New MLA variants/support report and usage guidance.
docs/moe_variants_guide.md	New MOE variants guide including routing/quantization/backend notes.
docs/gemm_variants_guide.md	New GEMM variants/tuning guide including configs and backends.
docs/quantization_guide.md	New quantization guide (FP8/MXFP4/INT4, fused ops, KV cache).
docs/normalization_guide.md	New normalization guide (RMSNorm/LayerNorm/GroupNorm + fusions).
docs/rope_guide.md	New RoPE guide (formats, backends, env selection, fused paths).
docs/kv_cache_guide.md	New KV-cache management guide (layouts, quantization, fused cache ops).
docs/elementwise_activation_guide.md	New elementwise/activation guide including fused quantization paths.
docs/sampling_guide.md	New sampling guide (greedy/random/mixed, top-k/top-p ops).
README.md	Updated installation instructions and added supported-operators index.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.