Recommodation for merge:

No.

There're some unit tests still needs to be added.

What it fixes

This PR is trying to fix #1921, along with other bugs that suffers the same reason.

The root cause is discussed in #1926, that when converting a torch.listconstruct, if any value in it depends on a dynamic shape which roots in the net input, the converted result is a bare python list and failed to parse in the subsequent op.

Design of this fix

This fix happends inside listconstruct parsing, runs a DFS from the current node to each parent layer all the way to the net root. And works on this asssumption:

If any value depends on a dynamic shape which roots in the net input, then there is an op gather, that gathers that dynamic shape from net input dims

What it does and does not

1. This fix does not affect the original branch for all compile-time scalar constants case
2. This fix still holds the original branch for all unexpected case
3. This fix addes another branch, only targets for an op gather, and what it gathers from is not a name in context.torch_graph.nodes, i.e. the net inputs.