✨[Feature] Support cases where LowerGraph() introduces input whose values are stored in Param

[bowang007]

Is your feature request related to a problem? Please describe.
#1024
There are some models that LowerGraph() might introduce new input to the entire graph after lowering, which may cause the missing value error. As what is happening here #922 (comment) and here: #922 (comment).

Describe the solution you'd like
For now, in most cases it fails at the shape analysis phase since Torch-TensorRT cannot find the value for the missing input. We could introduce a pass to modify the graph and substitute the input by some constant or we can just add some mechanism to use the values from Params directly.

In the future Torch-TensorRT could handle these cases.

[github-actions]

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

[ncomly-nvidia]

@bowang007 this seems like a bug as well. partitioning should never change the I/O of the graph. CC: @narendasan