RFC: Remove FAST_FLOAT, TFLOAT

[amitdo]

We should unconditionally use 32-bit float.

[stweil]

I already started experiments with even smaller float data types (like they are used in GPUs) because this would accelerate the training. Up to now my experiments were not successful, but who knows, this might change in the future with better compiler support and the right libraries. Therefore knowing the code locations and having a special data type is still very helpful.

But I also don't think that anybody still has the need for the old double implementation.

[amitdo]

But I also don't think that anybody still has the need for the old double implementation.

My suggestion:

• In cmake/autotools,FAST_FLOAT should always be defined (remove or comment out the televant config code).
• Remove all the double (64-bit) dot product code.