Fix problem with gradients accumulating

[AlekseySh]

The problem arises because a gradient accumulates in the layers with each new transformation.
Let's imagine, that original model works well with batch_size 128 (for example) and GPU was fully loaded, then wrapped model with several transform will crash for similar batch_size = 128, it will work only for batch size ~= 40.
To fix this problem I added torch._no_grad().

PS. If you now about this problem and assume that this code must be added in outer function, then you have to change your example snippet from readme because it doesn't include no_grad(). But in my opinion, it will be better left this code here, inside the forward.

[qubvel]

Hi, thanks for proposal.
In my opinion, despite it is called "test" time augmentation wrapper, it is still possible to use it for training. So, it is better to add a note about no_grad().