fix(quantization): Inherit Fields should not copy into empty provisional qspec

[guru-desh]

While I was quantizing a model with graph mode set to True, I ran into this issue:

(shortened traceback to only focus on important parts)

    def prepare(
        self,
        example_inputs: tuple[Any, ...],
        dynamic_shapes: dict[str, Any] | tuple[Any] | list[Any] | None = None,
        export_with_no_grad: bool = True,
    ) -> torch.fx.GraphModule:
	...
        try:
>           prepared_model = prepare_qat_pt2e(exported_model, quantizer)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
...
.venv/lib/python3.11/site-packages/coreai_opt/quantization/_graph/_qspec_resolution.py:145: in _assign_group_specs
    concrete = _build_concrete_spec(group.qspec)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
...

    def _build_concrete_spec(qspec: ProvisionalQSpec) -> TorchAOQuantizationSpec | None:
...
        missing = [
            field_name.name for field_name in _SPEC_KWARG_FROM_FIELD if field_name not in qspec.fields
        ]
        if missing or FieldName.QUANTIZATION_TARGET not in qspec.fields:
            # Every observed slot is seeded from a whole QuantizationSpec, so a
            # partial group is a generation bug. Don't paper over it with defaults.
>           raise ReconciliationError(
                f"Cannot rebuild a QuantizationSpec: reconciled group is missing "
                f"{missing or ['QUANTIZATION_TARGET']}. Present: "
                f"{sorted(f.name for f in qspec.fields)}."
            )
E           coreai_opt.quantization._graph._qspec_types.ReconciliationError: Cannot rebuild a QuantizationSpec: reconciled group is missing ['DTYPE', 'QFORMULATION', 'GRANULARITY', 'FAKE_QUANTIZE_CLS', 'QPARAM_CALCULATOR_CLS', 'RANGE_CALCULATOR_CLS', 'SCALE_DTYPE']. Present: ['FLOAT_RANGE', 'QSCHEME'].
.venv/lib/python3.11/site-packages/coreai_opt/quantization/_graph/_qspec_resolution.py:204: ReconciliationError

This was occuring in a model that had an attention layer and did the following operations as part of that attention module:

linear -> reshape -> permute

In the first part of annotation reconciliation, we traverse the entire graph. Here, we make two provisional qspecs:

1. Linear gets a filled provisional qspec as it was targeted in the user config
2. permute gets a provisional qspec as ShareObserverInstance creates one as its rank preserving. This provisional qspec covers reshape's output + permute's input + output

Because permute touches reshape, this part of the annotation reconciliation loop contains the reshape node. This then creates an InheritFields constraint. InheritFields apply then triggers which copies linear's qscheme and float range into permute's provisional qspec, which is the bug. It should not have copied linear's qscheme + float range as the provisional qspec was empty, so it creates a malformed provisional qspec triggering the error.

The fix is one line which is to add target_qspec.fields into the if statement so .apply() does nothing

[anotheranshu]

The proposed change to the setter is optional, but can you add a test which runs prepare on a Linear -> Reshape -> Permute model?