torch: fix compression when using apex.amp

[pleasantrabbit]

• use the correct ctx when decompressing data
• use tensor.copy if tensor.set fails. apex.amp monkey-patches optimizer.zerograd to set all grads to None. torch creates a new p.grad for each parameter. calling p.grad.set() causes an error:

  RuntimeError: set_storage is not allowed on a Tensor created from .data or .detach().
  If your intent is to change the metadata of a Tensor (such as sizes / strides / storage / storage_offset)
  without autograd tracking the change, remove the .data / .detach() call and wrap the change in a `with torch.no_grad():` block.
  For example, change:
  x.data.set_(y)
  to:
  with torch.no_grad():
      x.set_(y)

Signed-off-by: yulu.jia yulu.jia@bytedance.com