Fixed nn.functional.pad constant mode

[vfdev-5]

Description:

• Fixed nn.functional.pad constant mode
• Updated tests

The issue was with the fact that UNARY_POINTWISE does not properly handle bdim > 0. The fix is to use a macro which explicitly calls moveBatchDimToFront.

Related to #240

[vfdev-5]

@zou3519 please review this PR as well. Thanks

[Chillee]

Sorry, I don't understand why UNARY_POINTWISE doesn't support bdim > 0 🤔

Oh, is it specifically that for constant_pad_nd it screws up the padding argument?

But LGTM and passes the tests :)

[zou3519]

Sorry, I don't understand why UNARY_POINTWISE doesn't support bdim > 0 🤔

Oh, is it specifically that for constant_pad_nd it screws up the padding argument?

yes. The padding argument is a list of pads for the last dimensions. For example, constant_pad_nd(x: Tensor[2, 3], pad=[2, 3, 4, 5]) would pad dimension 0 using (2, 3) and dimension 1 using (4, 5).

Now, if your batch dimension is in the middle (e.g. constant_pad_nd(x: Tensor[2, B, 3], pad=[2, 3, 4, 5])) then we need to first move it to the front so it doesn't get padded by mistake