Add initial support for joint sparsity pipeline

[u-simha]

What:

• Add support for testing out sparsity with quantization & palettization
• This is mostly for verifying export support for joint sparsity

How:

• Add a private _sparsity field to the QuantizationSpec and PalettizationSpec to sparsity the weights before the fake quantization & fake palettization is applied
• Update prepare_for_export for both quantization & palettization to insert the sparse_to_dense ops in the correct order
• Add some basic end-to-end testing to validate the ops are inserted and the SNR of the prepared model and the exported model match

Constraints:

• Supported for post-training quantization & post-training palettization (we use the mask computed in the prepare() call)
• For quantization, only symmetric and integer-based quantization is supported; Only graph mode is supported
• For palettization, only per-tensor granularity is supported; quantized LUT and per-channel scales are not supported

Example quant + sparsity config:

config = QuantizerConfig(
            global_config=ModuleQuantizerConfig(
                op_state_spec={
                    "weight": QuantizationSpec(
                        dtype=torch.int8,
                        qscheme=QuantizationScheme.SYMMETRIC,
                        granularity=PerTensorGranularity(),
                        _sparsity=0.5,
                    )
                },
                op_input_spec=None,
                op_output_spec=None,
            ),
            execution_mode=ExecutionMode.GRAPH,
        )

Example palett + sparsity config:

config = KMeansPalettizerConfig(
            global_config=ModuleKMeansPalettizerConfig(
                op_state_spec={"weight": PalettizationSpec(n_bits=8, _sparsity=0.5)}
            )
        )