Use mmap for loading models

[snunez1]

The system currently loads the entire model into memory, which limits the size to available RAM. By using mmap we can load larger models. This is also a good chance to rework read-checkpoint, which is written using binary-types. In hindsight, binary-types wasn't a good choice for loading checkpoints because there is a lot of variability in the binary format. See make-vocabulary for an example of using mmap.