fix tensor_packbits_cuda_kernel for when batch size is not a multiple of bit count

[mahdizaferanchi]

Fixed this issue.

The new kernel also makes sure that padding bits are all zeros. Although I think this isn't necessary for correct operation.

Also added __repr__ to PackBitsTensor which helps detect the problem or the lack of it and is also generally useful.

[Felix-Petersen]

Thanks @mahdizaferanchi for fixing this bug! 🪲