Make the appropriate backend DimensionNode visible to LTC core

[miladm]

🚀 The feature, motivation and pitch

With the current implementation of DimensionNode in LTC, each backend has its own implementation (PyTorch/XLA, TorchScript).

At the same time, shape inference builds off of LTC core classes leading to this PR failing to build shape inference implementation for expand.SymInt. Please make the appropriate backend DimensionNode visible to LTC core.

Solution alternatives based on offline chat with @Krovatkin.

• IR core to access the correct backend implementation
• Use multiple inheritance

CC @wconstab @JackCaoG

Alternatives

No response

Additional context

No response

[wconstab]

Can we leverage the non-native IR codegen (#76535) to solve this, by making DimensionNode one of the generated IRs that backends own?

I guess that is orthogonal. That would be equivalent to the status quo, but save some code-duplication.

We could expand on the IR Builder interface to let backends build their own DimensionNode through a core/backend API.