Fixes for initializers/convaware.py

[torzdf]

This PR fixes 3 issues in initializers/convaware.py:

• Fixes the transpose on rank == 4 to correctly return column and row in the correct order (re: Issue ConvolutionAware error #479)
• Respect passed in dtype to __call__ (as per it's parent class) rather than relying on K.floatx()
• Correctly return a K.variable() of the correct dtype rather than a numpy array of dtype K.floatx()

Verification

• This PR can be verified by initializing a conv2D layer with the ConvolutionAware initializer. As per issue ConvolutionAware error #479, it will fail if run from the old code but succeed with this PR.
• Additionally, if wrapping this code in another initializer (for example an ICNR initializer) with a different dtype to that stored in K.floatx(), this PR will succeed where the original code will fail.
• Finally, the original code will fail if wrapping this initializer inside another initializer (eg. ICNR) using Tensorflow backend as the returned numpy array forces a tensorflow Channel Order error. By returning a K.variable() from this initializer, that error is avoided.

---

This pull request fixes #479