Support negative pad values

[codeislife99]

Negative pad values error out currently. This PR removes it and handles the error correctly for negative pad values.

[codeislife99]

@anijain2305 @comaniac @zhiics @trevor-m PTAL

[anijain2305]

Please add a unit test

[trevor-m]

Thanks @codeislife99 ! It looks good to me, but could you also add a test case in test_op_level2.py?

[codeislife99]

Added relevant test cases.

[codeislife99]

@anijain2305 @trevor-m
Please re-review.
One dilemna here is that when shape is dynamic I cannot check if the existing shape + padding would be greater than zero in infertype. This will lead to a memory error during executing when it wont be able to create memory for negative tensor sizes. And the error would be quite bad because the developer/user would not have a clue where it is coming from in a big model.

Does anyone have any idea of how to solve this ? Should we display a warning whenever we encounter negative padding ?

[comaniac]

When handling the dynamic shape, you need to improve the shape function to check if the shape after padding is still legal on the fly. If it cannot be simply covered in this PR, we can focus on the static shape first and still throw out errors for negative padding values when the shape is dynamic.

[comaniac]

LGTM

[anijain2305]

LGTM

[anijain2305]

Thanks @codeislife99 @comaniac @trevor-m This is merged