Avoid mandatory torchao dependency for quantized_decomposed registration

[npow]

Summary

#290 pointed PyTorch >= 2.11 at torchao.quantization.pt2e.quantize_pt2e after torch.ao.quantization.quantize_pt2e was removed upstream (confirmed absent on 2.13.0+rocm7.2). That works, but makes torchao a hard new dependency for what is only a registration side effect here (populating torch.ops.quantized_decomposed).

torch.ao.quantization.fx._decomposed provides the exact same registration with zero extra dependencies — confirmed live: torch.ops.quantized_decomposed.quantize_per_tensor is absent before this import and present after, on 2.13.0+rocm7.2 (real MI300X). It's also where quantize_pt2e itself sourced this registration from on older PyTorch, so it's a strictly more fundamental import rather than a workaround.

Practical motivation: torchao ships CUDA-specific compiled extensions (_C_mxfp8, _C_cutlass_90a) that fail to load on a ROCm build with no functional impact — confusing/alarming noise for anyone who doesn't already depend on torchao for something else.

This PR prefers the dependency-free path and falls back to the existing #290 behavior (torchao on >=2.11, in-tree quantize_pt2e below that) only if _decomposed is ever unavailable, so it's strictly additive and preserves existing behavior everywhere the new path doesn't apply.

Test plan

• Real MI300X hardware (ROCm 7.2.4, torch 2.13.0+rocm7.2): clean import torch_migraphx with no torchao installed
• torch.compile(backend="migraphx") compiles and runs correctly on a transformer-block-shaped module: cosine similarity 1.0 vs eager

🤖 Generated with Claude Code