Add op(roll) | feat(torchlib)

[xiaowuhu]

In _roll_input_wrangler:

1. convert shifts from int to tensor
2. convert dims from int or tensor to attributes list, so we can use loop in trace_only mode to process aten_roll(x, (y1, y2), (z1, z2)) case.

[codecov]

Codecov Report

Merging #905 (c39595c) into main (7402a96) will increase coverage by 0.07%.
The diff coverage is 92.30%.

@@            Coverage Diff             @@
##             main     #905      +/-   ##
==========================================
+ Coverage   76.84%   76.91%   +0.07%     
==========================================
  Files         112      112              
  Lines       13619    13669      +50     
  Branches     1386     1397      +11     
==========================================
+ Hits        10466    10514      +48     
+ Misses       2805     2804       -1     
- Partials      348      351       +3     

Files Changed	Coverage Δ
...ipt/tests/function_libs/torch_lib/ops_test_data.py	96.20% <84.61%> (-0.68%)	⬇️
onnxscript/function_libs/torch_lib/ops/core.py	77.72% <94.87%> (+0.28%)	⬆️

... and 1 file with indirect coverage changes

[justinchuby]

dims is a Sequence that is by default empty: int[1] dims=[]. It will not be None, or an int.

[justinchuby]

LGTM. The last batch of comments mostly around updating the signature/logic to match the aten signature.

[justinchuby]

I also tried https://github.com/microsoft/onnxscript/pull/918/files only to realize concat doesn't take a dynamic axis

[justinchuby]

Thank you!