Apply fp16_op_block_list / fp16_keep_io_types to QDQ path (and vice versa)

[DingmaomaoBJTU]

Context

From PR #872 review feedback by @xieofxie:

consider also apply to qdq and vice versa
a trick is get all types from input model to get the reverse list

Currently fp16_keep_io_types and fp16_op_block_list in WinMLQuantizationConfig are only used when mode="fp16". The suggestion is to:

1. Allow FP16 conversion settings (like op_block_list) to also apply when running QDQ quantization (e.g., to skip certain ops from quantization).
2. Allow QDQ-specific settings to inform FP16 conversion (e.g., auto-derive the block list from the model's op types).

Proposed Approach

• Inspect all op types in the input model to auto-generate a sensible block/allow list
• Share relevant config fields across FP16 and QDQ paths where applicable
• Keep backward compatibility (explicit user settings always take priority)

Related

• PR feat: precision-driven quantization (FP16, RTN int4, static QDQ) via --precision flag #872 (precision-driven quantization)
• src/winml/modelkit/quant/config.py — WinMLQuantizationConfig