Check the scan op input for requires_grad

[haifeng-jin]

Resolves #8783

Added a test for LongTensor inputs, which would fail without this PR.
Only set carry.requires_grad to True when the dtype is floating point.

Return None, None in the backward() function if none of the outputs has gradients.

Questions:

1. Why don't we just set carry.requires_grad the same as the user input init? (I did run into some errors when I tried that)
2. Do we also need to return None for init.grad in backward(), when only x requires grad?
3. It fails when init is LongTensor and x is float32 because it requires carry.dtype == init.dtype, while carry is float because it is produced by int + float. Do we need to support this case?