Fix statistics of the residual image

[AlexKurek]

First:
resid_gaus = cp(img.ch0_arr)

Later the algorithm checks if the mask is defined. If it is, it injects "new" NaNs into resid_gaus

if isinstance(img.mask_arr, N.ndarray):
    pix_masked = N.where(img.mask_arr == True)
    model_gaus[pix_masked] = N.nan
    resid_gaus[pix_masked] = N.nan (1)

The algorithm tries to isolate the non-Nan pixels:
resid_gaus_non_masked = resid_gaus[N.where(~N.isnan(img.ch0_arr))
The problem is, that it searches for the pixels that are not NaN in the original image. But "other" NaNs were added to resid_gaus in (1). So resid_gaus_non_masked still may have NaNs. In that case later N.mean(), N.std() etc. will return NaN-s for the entire array.

EDIT:
Reverted Same for shapelets since it crashes the test. Maybe lets stick to resid_gaus in this PR.

[AlexKurek]

The test is failing for unrelated reasons ( #309 (comment) )