[Frontend] Add quantisation pass

[stomfaig]

This PR is implementing a new pass for quantising graphs. We repeatedly evaluate each node, check if we can decide if it can be quantised, try to quantise it, and repeat the process. For more details see the code, it is extensively documented.

Note: the code also contains the code from #641, so that it is runnable locally. This poc only supports simple models like below, and is intended as a demonstration of the approach. The changes strictly associated with this pr are only in frontend/Python/graph/transform/quantise.py

class SimpleNet(nn.Module):
    def __init__(self):
      super().__init__()
      self.linear = torch.nn.Linear(28, 28)

    def forward(self, x):
        return self.linear(x)

Test driver to see the changes:

def main():

    model = SimpleNet()

    dynamo_compiler = DynamoCompiler(
        primary_registry=tosa.ops_registry,
        aot_autograd_decomposition=inductor_decomp,
        func_name="simple",
    )

    graphs = dynamo_compiler.importer(
        model,
        x=torch.rand((1, 1, 28, 28)),
    )

    assert len(graphs) == 1

    for graph in graphs:
        graph.init_op_group()

    graphs[0].perform(
        [quantise_graph]
    )

    g_body = graphs[0]._body
    for op in g_body:
        print("\n---")
        print(f"{op._name} : {type(op)}")
        print(f"children: {op._children}")
        print(f"parants: {op._parents}")
        print(f"tensor_meta: {op._tensor_meta}")

Depends on #641, #644

[stomfaig]

cc: @R-Tars

[zhanghb97]

Looks like progress is going well. Just a reminder: we upgraded to LLVM 22. Don't forget to rebase to avoid potential compatibility issues.