Missing pooling options, inconsistencies with CuArrays

[denizyuret]

There are two mean pooling operators in cudnn, one that includes and one that excludes the padded values. The current meanpool operation only supports including the padded values, we should add the other:

https://docs.nvidia.com/deeplearning/sdk/cudnn-api/index.html#cudnnPoolingMode_t

[denizyuret]

There is an option, maxpoolingNanOpt, for not propagating NaN values during max pooling in CUDNN:

typedef enum {
    CUDNN_NOT_PROPAGATE_NAN = 0,
    CUDNN_PROPAGATE_NAN     = 1,
} cudnnNanPropagation_t;

We do not have an equivalent mechanism in NNlib. The current NNlib behavior propagates NaNs, so it is equivalent to option 1. It would be better to have the default be option 0.

[denizyuret]

Related also: CUDNN_POOLING_MAX excludes padding by default, this creates an inconsistency with CuArrays:

julia> x = randn(4,4,1,2)

julia> cx = CuArray(x)

julia> maxpool(x, PoolDims(x,2,padding=1))
3×3×1×2 Array{Float64,4}:
[:, :, 1, 1] =
 0.0       0.773338  0.0
 0.0       2.17463   0.751506
 0.299839  0.8081    0.489329

[:, :, 1, 2] =
 0.0      1.70753   0.103033
 1.26366  0.58986   0.905675
 0.0      0.927308  0.0

julia> maxpool(cx, PoolDims(x,2,padding=1))
3×3×1×2 CuArray{Float64,4}:
[:, :, 1, 1] =
 -0.976195  0.773338  -0.409254
 -0.231234  2.17463    0.751506
  0.299839  0.8081     0.489329

[:, :, 1, 2] =
 -0.119529  1.70753    0.103033
  1.26366   0.58986    0.905675
 -0.2345    0.927308  -0.978941

[CarloLucibello]

the inconsistency has been fixed in #238, leaving this open for the missing pooling optinos