[PyTorch] torch.listconstruct causing issue for other ops

This is the root cause to many issues. When symbolic shape is involved in `torch.listconstruct`, instead of a CoreML tensor, we simply return the list as is
```
def _array_construct(context, node, array_type):

...

else:
# If at least one input to the construct op is non-const, collect
# the inputs and add them directly to the context. Ops that use this
# node's output will take the list directly as input.
context.add(array_type(inputs), node.name)
```

## Appendix 1: Issues Sharing the Same Root Cause
* [#1303](https://github.com/apple/coremltools/issues/1303)
* [#1921](https://github.com/apple/coremltools/issues/1921)
* [#1991](https://github.com/apple/coremltools/issues/1991)
* [#2037](https://github.com/apple/coremltools/issues/2037)

## Appendix 2: Ops Impacted by the Root Cause
* `torch.GroupNorm`
* `torch.pad`
* `torch.index_put`

Start at the _array_construct entry point shown in the issue and trace how symbolic-shape torch.listconstruct outputs are consumed by GroupNorm, torch.pad, and torch.index_put conversion. Review the related issues for expected behavior and add regression coverage for the affected operations. Done means those operations no longer fail because a symbolic construct output is passed through as a raw list.

Written by the indexing model from the issue text.