Featurewise zero center and featurewise stdnorm fix

[ziky90]

This PR fixes featurewise zero centring and normalization in order to work also with datasets that are not float32 eg. uint8 images.

It also replaces redundant for loop by numpy broadcasting.

[jakubsimanek]

This indeed fixes the issue where batch[i] has incompatible dtype with the self.global_mean.value

When batch[i].dtype is uint8 and self.global_mean.value.dtype is float64 the old implementation of _featurewise_zero_center results in:

   batch[i] -= self.global_mean.value
TypeError: Cannot cast ufunc subtract output from dtype('float64') to dtype('uint8') with casting rule 'same_kind'

[ziky90]

@aymericdamien is there any progress in reviewing this PR.
It attempts to fix quite serious bug in the tflearn/data_preprocessing.py. As @jakubsimanek writes caused by that numpy can't cast arrays using -= operation.

[arundasan91]

@ziky90 , This works. Thanks! 👍