Expose --quantize_dtype CLI arg to control dtype during quantization

[rhn19]

Fixes #226

Problem

The model is loaded in --dtype before quantization. There is no way to quantize in a different dtype (e.g. to match QAT training precision).

Changes

• Add --quantize_dtype CLI argument (choices=["float32", "float16", "bfloat16"])
• Wire it through all task loaders (causal_lm, masked_lm, asr, multimodal_text_to_text) to quantize_model_()
• In quantize_model_(): cast to quantize_dtype before quantization, restore original dtype before unwrap_tensor_subclass

The restore-before-unwrap ordering matters: unwrap_tensor_subclass breaks .to() for quantized tensor subclasses, so the dtype cast must happen while the subclass wrappers are still intact.

Usage

optimum-cli export executorch \
   --model merged-qat-model/ \
   --task text-generation --recipe xnnpack \
   --qlinear 8da4w --quantize_dtype bfloat16 \
   --dtype float32 \
   --output_dir output/

No behavior change when --quantize_dtype is not specified.