Llama phase split quantisation support

[YuxuanHan0326]

PR Summary (vs main)

1. Phase-split config for the Llama quantized path

• Quant configs can be defined separately for prefill and decode.
• Runtime phase is inferred per decoder-layer forward and propagated via context, so modules run with the correct phase config.

2. Decode behavior is explicitly controlled

• decode_policy="fp_only" keeps decode in FP-compatible mode.
• decode_policy="quantized" enables decode quantization using decode sub-config.
• A fail-fast guard prevents mixed decode policies across quantized Llama modules.

3. Core quantized Llama modules are phase-aware

• attention, mlp, rms_norm, and MX linear modules now consume runtime phase + per-phase config at inference time.
• Supports practical modes: prefill-only quant, decode-only quant, and different prefill/decode configs.

4. MX linear decode bank path is integrated

• LinearMXFP/LinearMXInt select decode-path banks based on phase/policy.
• Decode bank refresh remains compatible with GPTQ + module replacement flow.

5. Compatibility and stability

• Legacy flat quant configs are still accepted (internally normalized to phase structure).
• Rotate symbol exports were restored to preserve existing import/registry contracts.
• Targeted split-phase tests pass, and real Llama+GPTQ chain checks produced readable outputs across all three phase combinations.

6. Scope boundary (not changed in this PR)

• Llama-focused only (no extension to other model families).
• No runtime re-quantization feature so currently only working for small models.
• No BFCL/evaluation pipeline changes.
• Existing non-phase quant workflows are intentionally left unchanged.