[Insight-developers] MI changes

[Lydia Ng]

As requested I have updated the MutualInformationMetric to
handle images with long tails using the minimum value method
as per discussion on tcon.

This change seems to impact on the choice of  the learning rate:
a larger value seems to be required to get the same convergence
rate.

I have also added a heuristic to test if the Parzen
window width is too small. An exception is thrown
if at least half the samples in B do not lie within
a Parzen window width of samples in A.

Cheers,
Lydia