Quantize before import and conversion to Linalg during CompiledModule.__new__()

[renxida]

Now, passes that need to happen on torch IR before conversion to linalg can be supplied as pre_import_passes to the constructor of StateUpdateModule or other classes that inherit from CompiledModule.

E.g.

    pre_import_passes = []
    if quantization == "int4" and not compile_to == "linalg":
        from shark_turbine.transforms.quantization import mm_group_quant
        pre_import_passes.append(mm_group_quant.MMGroupQuantRewriterPass)
    inst = StateUpdateModule(context=Context(), import_to=import_to, pre_import_passes=pre_import_passes)

[renxida]

This fixes #183

[IanNod]

Please run black on the files changed