✨[Feature] Exclude key functions from tracing during AOT

[gs-olive]

Context

As of #1921, we can exclude modules from acceleration in AOT using the global module-level registry. Still, there are certain functions which are also decomposed as a result of aot_module_simplified:

TensorRT/py/torch_tensorrt/dynamo/backend/backends.py

Lines 50 to 55 in e109049

	return aot_module_simplified(
	gm,
	sample_inputs,
	fw_compiler=make_boxed_compiler(custom_backend),
	decompositions=get_decompositions(),
	)

These functions include torch.functional.einsum, which is decomposed as follows:

##### THE MODEL
class SampleEinsum(torch.nn.Module):
    def forward(self, x):
        return torch.einsum('ii', x)

##### BEFORE AOT
graph():
    %l_x_ : torch.Tensor [#users=1] = placeholder[target=L_x_]
    %einsum : [#users=1] = call_function[target=torch.functional.einsum](args = (ii, %l_x_), kwargs = {})
    return (einsum,)

##### AFTER AOT
graph():
    %arg0_1 : [#users=1] = placeholder[target=arg0_1]
    %as_strided : [#users=1] = call_function[target=torch.ops.aten.as_strided.default](args = (%arg0_1, [4], [5], 0), kwargs = {})
    %permute : [#users=1] = call_function[target=torch.ops.aten.permute.default](args = (%as_strided, [0]), kwargs = {})
    %permute_1 : [#users=1] = call_function[target=torch.ops.aten.permute.default](args = (%permute, [0]), kwargs = {})
    %sum_1 : [#users=1] = call_function[target=torch.ops.aten.sum.dim_IntList](args = (%permute_1, [0]), kwargs = {})
    return (sum_1,)

Feature Proposal

Augment the module-level replacement registry with a function-level replacement utility to replace TRT-convertible functions like einsum and exclude those from AOT tracing.

[github-actions]

This issue has not seen activity for 90 days, Remove stale label or comment or this will be closed in 10 days

[gs-olive]

Completed in #1980