Fix import for torch>=2.12: quantize_pt2e removed

[beausifutmez]

Fix import torch_migraphx on torch >= 2.12 (quantize_pt2e removed)

Problem

import torch_migraphx fails for all users on recent PyTorch, not only those doing quantization:

File ".../fx/converters/quant_ops_converters.py", line 49, in <module>
    import torch.ao.quantization.quantize_pt2e
ModuleNotFoundError: No module named 'torch.ao.quantization.quantize_pt2e'

The module was deleted in pytorch/pytorch#169151
("Delete pt2e flow… migrated to torchao", merged 0e38e02), which ships in the torch 2.9+ line —
matching the report in #279 (torch 2.9.1) and reproduced here on torch 2.12.1.

Note the import exists purely for a side effect: registering the torch.ops.quantized_decomposed
ops used by the converter decorators just below it. Guarding the import alone is not enough — the
module then fails at decorator time with
'_OpNamespace' 'quantized_decomposed' object has no attribute 'quantize_per_tensor'.

Fix

The two layers PyTorch split apart matter here:

• The pt2e flow (prepare_pt2e/convert_pt2e) moved to torchao — that's what #169151 deleted.
• The quantized_decomposed ops remain in torch core, registered by
  torch.ao.quantization.fx._decomposed.

This import only needs the latter, so pointing it at fx._decomposed is the canonical, in-core,
zero-dependency registrar (not a workaround). Guarded for back-compat with torch < 2.12:

# Import required to populate torch.ops.quantized_decomposed.
try:
    import torch.ao.quantization.quantize_pt2e        # torch < 2.12
except ModuleNotFoundError:
    import torch.ao.quantization.fx._decomposed       # torch >= 2.12

One file changed; no behavioral change on older torch (it takes the try path unchanged).

Verification

Verified on torch 2.12.1+rocm7.2 (HIP 7.2), ROCm 7.2.4, Radeon RX 7900 XTX (gfx1100), Linux:

• import torch_migraphx succeeds; the migraphx dynamo backend loads.
• torch.ops.quantized_decomposed is populated (quantize_per_tensor etc. present).
• torch.compile(model, backend="migraphx") runs a conv+linear model and ResNet-50, output matching
  eager to fp32 epsilon (max |Δ| ≈ 1.2e-7).

fx._decomposed has existed across the affected range, so the same guard covers the torch 2.9.1
case reported in #279 (Windows / gfx1151) as well as torch 2.12 here.

Scope

This restores import + the existing quantized_decomposed converters only. Fully migrating
torch_migraphx's pt2e quantization flow to torchao is a separate, larger change and out of scope here.

Fixes #279