What does this PR do?

Fixes loading of pre-quantized ModelOpt models via from_pretrained(..., device_map=...),
which currently crashes with:

NotImplementedError: Cannot copy out of meta tensor; no data! ...

Note: this is required for the pre-quantized path to work at all with HF checkpointing enabled —
no combination of device_map / .to(device) load options works around it without this change.

Root cause

For pre-quantized models, NVIDIAModelOptQuantizer.check_if_quantized_param returned True
for every tensor. That routes the quantizer buffers ModelOpt recreates on restore
(_amax, _scale) through create_quantized_param, which assigns them into
module._parameters. Those buffers are then never materialized by the low-memory loader and
stay on the meta device, so the later device move fails.

Fix

• check_if_quantized_param: for pre-quantized models, only claim real parameters
  (return tensor_name in module._parameters); buffers fall back to the standard loader, which
  materializes _amax/_scale on their target device.
• create_quantized_param: guard the pre-quantized branch so a non-parameter reaching it
  raises clearly instead of silently mis-assigning.

Regression test

Added test_prequantized_serialization_with_device_map, which:

• Enables enable_huggingface_checkpointing() before quantization, save, and reload.
• Verifies modelopt_state.pth is produced.
• Reloads the pre-quantized checkpoint with device_map.
• Confirms ModelOpt _amax/_scale buffers are restored.
• Asserts no parameters or buffers remain on meta.

Validation:

RUN_NIGHTLY=1 python -m pytest \
  tests/quantization/modelopt/test_modelopt.py::SanaTransformerFP8WeightsTest::test_prequantized_serialization_with_device_map \
  -v -s

• Without this fix: fails with NotImplementedError: Cannot copy out of meta tensor; no data!
• With this fix: passes

Environment

Single ModelOpt version on both export and load (ModelOpt state is version-sensitive, so
producer and consumer must match):

• diffusers 0.40.0.dev0 (this branch)
• nvidia-modelopt 0.44.0 (quantize + load)
• torch 2.9.0a0+145a3a7bda.nv25.10 (NVIDIA NGC container build)

Who can review?

@sayakpaul