Implement torchscript as a graph backend

[justinchuby]

To use onnxscript APIs for constructing graphs in pytorch, we would like to create a version that uses torchscript as the graph IR. To do this, we probably need

1. Implement an Evaluator that calls g.op to add ops and functions as nodes
2. Implement a Tensor that wraps around the torch.Values

[titaiwangms]

1/11 meeting log:

1. Use context.manager to execute evaluator switch
2. implement kwargs adapter
3. Demo on multiple ops model (MUST use atenlib, which is ONNXFunction)

Future consideration:

1. Tensor adapter
2. Build the graph from scratch with onnxscript API

[gramalingam]

Just noting down a question here, hopefully our meeting later today will answer that: is the intention to have:

(i) a way of building torch IR using pythonic syntax (just like onnxscript), or
(ii) To have a deeper integration with onnxscript?

I am guessing it should be (i). In particular, do we ever expect core onnxscript to call the torch evaluator? I am assuming not.

[justinchuby]

@titaiwangms for now, an aten function can be called like

from onnxscript.function_libs.torch_aten import ops

ops.core.aten_add()

[titaiwangms]

@titaiwangms for now, an aten function can be called like

from onnxscript.function_libs.torch_aten import ops

ops.core.aten_add()

ops doesn't have module core, could you check on that? I am importing
from onnxscript.function_libs.torch_aten.ops import core