Cascade2d - with stride > 0 the result is wrong

[blasscoc]

If we decimate at each convolution, we need to also decimate the wavelets for the next order in the filter bank, otherwise the spectrum will be all messed up.

[blasscoc]

Go ahead and design an object that will implement a "downsampling" by the correct factors, at each step of the transform.

The scale of the finest wavelet will decimate by a factor of 2, then next by 4, etc.

The convolution with the father wavelet will decimate by a factor of 2^J.

Differ the actual implementation of the decimation scheme to a third object or function so we have some flexibility to develop.

[blasscoc]

From Kymatio; subsampling for convolution, signal with father wavelet: U_1_c = subsample_fourier(cdgmm(U_0_c, phi[0]), k=2**J)

Subsampling after application of convolution with \psi: U_1_c = subsample_fourier(U_1_c, k=2 ** j1)

Subsequent subsampling after application of convolution with father wavelet: U_2_c = subsample_fourier(cdgmm(U_1_c, phi[j1]), k=2**(J-j1))

Parameters

    x : tensor_like
        input tensor with at least 5 dimensions, the last being the real
         and imaginary parts.
        Ideally, the last dimension should be a power of 2 to avoid errors.
    k : int
        integer such that x is subsampled by 2**k along the spatial variables.

[blasscoc]

After second order convolution: if(j1 < j2): U_2_c = subsample_fourier(cdgmm(U_1_c, psi[n2][j1]), k=2 ** (j2-j1))

Convolving with father wavelet, after second order convolution: U_2_c = subsample_fourier(cdgmm(U_2_c, phi[j2]), k=2 ** (J-j2))