Dynamic setting of input batch size and support for multiple nets in Matlab wrapper

[kencoken]

I'm proposing a pull request which introduces the following new functionality:

• The ability to override the first dimension of any input blob(s) by specifying the optional force_input_count parameter when initialising a network
• Extensions to the Matlab wrapper to allow multiple network instances to be maintained simultaneously (by keeping a hashmap of handles)

These are both prerequisites for an extended Matlab wrapper for mid-level feature computation using Caffe which I am planning to distribute soon, ported from the deepeval-encoder package, with the intention of replacing it with a new package which builds upon caffe.

Please let me know if there is anything I can do to assist in the integration of these changes to the main Caffe branch, or if there is a more appropriate way in which to port these changes. In particular, am aware that the force_input_count parameter is quite specific, but nonetheless opted to issue a PR given that, since it is an optional argument, it does not change the existing API if not used. It may be that I maintain these changes in a separate forked repo, but thought I'd issue a PR just in case these changes were useful to others.

[shelhamer]

Thanks for the pull request and apologies for the silence, but matcaffe has been updated altogether in #2505. The new interface includes multiple nets, reshaping, and more. If you see any improvements to make in the new interface, please do send a PR!

I hope the new interface is useful for your deepeval package.