* Add test revealing issue in current cudnn pooling operation. To fix we need op::MaxPoolBackprop to take the fprop output as an input argument.

* Formatting.

* Add comments and test description. Increase tensor size well past most cache sizes.

* exclude intelgpu