This makes a few changes, specifically to fasttext, but it can apply to other
DJL ml wrappers as well. First, it creates several passthrough utility classes
with the goal that ml models can be loaded through the model zoo and run through
the predictor.

Next, it modifies the base of the ml construct to better support engines that
have multiple applications. Each application now manifests as a type of
SymbolBlock rather than a model. Then, the single model class can run any of
them. The model is still created for each engine because it contains general
loading functionality that can determine which block should be used to load the
given target.

It also has to update the fasttext JNI to support this. First, it fixes the
modelType to actually return different results. Then, it modifies the
predictProba method to use an ArrayList instead of an Array. The main difference
is because it is possible to pass a topk of -1 in order to load all elements.
However, this doesn't work with the previous array setup.

Change-Id: I5fbaceb2ac4711a942b9d15dac6e6fed386dd46e

Change-Id: I7b03f8958594c01d3ab2f722dbb7a573c6d3109e

Change-Id: Iadf598d930dca954ee1a0f450012cd4b4b2baab5