Fix RuntimeError of loading model with external tensor

[jcwchen]

Description:

• Set the parent directory of model path to load external tensors
• (TODO) Resolve new error after fixing the previous one: (It will be done in another PR)

[ONNXRuntimeError] : 2 : INVALID_ARGUMENT : corrupted protobuf data: tensor shape size(1) does not match the data size(0) in prototensor shape size(1) does not match the data size(0) in proto

Motivation and Context
onnx/onnx#3157
Reload and save the onnx model with external data format will cause runtime error:

onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from chitchatexporttest/median_sampled_qgen_48_past_fp32/median_sampled_qgen_48_past_fp32.onnx failed:This is an invalid model. Error in Node:ConstantOfShape_11958 : Data of TensorProto ( tensor name: ConstantOfShape_11958_attr::value) should be stored in median_sampled_qgen_48_past_fp32.onnx.data, but it doesn't exist or is not accessible.

[jcwchen]

All updated. Thank you for the review and this PR is ready.

[snnn]

why it must be ended by '/'?

[jcwchen]

https://github.com/onnx/onnx/blob/bd2ff65b775580d21ab38cb655a04854c348ff49/onnx/checker.cc#L133
It is required here because the code in ONNX does not handle path join...

[snnn]

Then onnx should fix it! It is always safe to add an additional path separator there when you do path join. But, it seems a little bit wired to do it here without knowing the result will only be used in path join.

[snnn]

And instead of calling std::ifstream, it should use fstat to test if a file exists: https://linux.die.net/man/2/fstat

[jcwchen]

Makes sense. I am proposing a PR in ONNX to fix it before checking this PR in. Thank you for bringing this up.

[jcwchen]

Here is the PR to fix that in ONNX: onnx/onnx#3214

[jcwchen]

A path join fix has been merged in ONNX: onnx/onnx#3214. Then this PR would be easier. Set the model directory without giving the separator and update ONNX commit. This PR is ready for review. Thanks.