[ONNX] Simplify repeat_intereleave export for scalar-valued 'repeat' (#100575)
This PR simplifies the ONNX export of torch.repeat_interleave when 'repeat' is a scalar value (so each index in the input is repeated the same number of times). (Issue #100438)
Here is a before/after of a simple model export:
```python
# Model + export code
import torch
class RepeatInterleaveModel(torch.nn.Module):
def forward(self, x):
return x.repeat_interleave(2, dim=-1)
args = (torch.rand((2, 2, 16)),)
model = RepeatInterleaveModel()
torch.onnx.export(model, args, "repeat_interleave.onnx", opset_version=17)
```
**Before (static shapes)**
-----
**Before (dynamic shapes, second graph is Loop body)**
<p float="left">
<img src="https://user-images.githubusercontent.com/46343317/236029895-20b0ae0a-240f-466d-bb01-e619ec5967ad.png" width="45%" />
<img src="https://user-images.githubusercontent.com/46343317/236029915-e67b808a-029b-4997-bc05-1ce59eec409a.png" width="47%" />
</p>
-----
**After (for both static and dynamic shapes)**
<img src="https://user-images.githubusercontent.com/46343317/236015235-633811cb-09a2-435d-a293-1b2bcb7dea50.png" width="66%" />
-----
This PR also fixes a bug where the exporter throws an expection when the input has dynamic shapes and the 'dim' parameter is not specified to torch.repeat_interleave. Also adds a new testcase to cover this. (Issue #100429)
Fixes #100438 and #100429
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100575
Approved by: https://github.com/BowenBao
