Issue error when serializing ANN models without MLPACK_ENABLE_ANN_SERIALIZATION

[rcurtin]

This is a fix for @yurivict's suggestion in #3449.

If MLPACK_ENABLE_ANN_SERIALIZATION is not defined, then we issue a runtime exception. I tried to make it a static assertion only, so the failure was at compile time, but I couldn't quickly find a way to make the compilation fail only when FFN::serialize() was called (or RNN::serialize()). I think the runtime error is fine though.

I left an "escape mechanism": if the user defines MLPACK_ANN_IGNORE_SERIALIZATION_WARNING, then the exception is not thrown. This is in case a user wants to specifically instantiate serialization for only certain layers, which would be done with custom CEREAL_REGISTER_TYPE() calls.

[zoq]

Looks like this is causing some issues with the test cases.

[rcurtin]

I think it actually has to do with cotire and precompiled headers. When compiling simple test programs by hand with and without MLPACK_ENABLE_ANN_SERIALIZATION, everything seems to work as expected. I meant to dig into it today, but instead spent my time debugging sorting on the GPU... maybe tomorrow? :)

[rcurtin]

Ok, got it figured out now. The fix was relatively straightforward and had to do with cotire, like I thought.

[zoq]

Nice