Phase 2: replace dict-style quant_config access with LayerQuantSpec a…

[thpereir]

…cross all models and ops

Motivation

Following up DeepSeek R1 enablement with per layer quantization support this converts all other models to use the same structured quantization config instead of a dict.

This PR depends on both #236 and #268

Technical Details

1. Clean type boundary — LayerQuantSpec is a frozen dataclass; impossible to accidentally mutate or misspell fields.
2. Single resolution API — resolve(prefix) handles exclude-list, per-layer overrides, pattern matching, and global fallback in one call with documented priority. Less error prone than previous approach of accessing the dict. What was quant_config.get_layer_quant_config(f"{prefix}.fused_qkv_a_proj") becomes quant_config.resolve(prefix)
3. Parser registry — @register_quant_parser("quark") means new quantization formats are self-contained modules. In the current approach adding a new quantization method means editing QuantizationConfig.__init__ directly.

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.