Maxpool bwd

[rocking5566]

Maxpool backward is dx = dy[index], which is same as Put operation from

https://pytorch.org/docs/stable/generated/torch.Tensor.put_.html
https://numpy.org/doc/stable/reference/generated/numpy.put.html

Hence, I implemented Put kernel and uses it to implement Maxpool backward.
However, if each sliding windows will be overlap (eg. window size = 3, window stride = 1), we need to use atomicAdd() to add the gradient. In this case, Maxppol backward is dx += dy[index]
So, my Put kernel let the user to specify the memory operation different than the usually used setting the value.
MemOp(Dx, Dy, index)

However, fp16x1 and bf16 does not support atomicAdd. In this case, Put kernel will output fp32 and another casting kernel is used to cast fp32 to fp16 and bf16.

Note: This P.R assumes all the input and output tensors are in packed memory space.

[rocking5566]

Similar to DeviceElementwiseImpl,
We need to get grid size until we solve #266
Hence, I hardcode the grid size for now.

[qianfengz]

For referene_pool_fwd.hpp and the Class name, suggest to include "nhwc", as the codes insides shows the supported layout is "NHWC"