Improve stability and composability

[amreis]

Hello!

Thank you for coding this nice implementation of KDE in Torch. I've been using it, but have run into a few issues.

• Sometimes, the log-likelihood is -inf, which can cause downstream problems when computing gradients (they become NaN). I've added an option for a user to specify an eps value that gets added to the density before the log is taken. Using a small positive value improves stability there.
• Using this KDE as part of another module in Torch used to run into some issues because KernelDensity never calls super().init(). This means it does not populate some variables (self._modules, for instance) that Torch uses internally. I've added the call.

I've taken the liberty of removing some trailing whitespace too.

Thank you once more for the package =)

[rudolfwilliam]

Thank you for the contribution. This is indeed a useful feature.