GH-1469: Add command-line tools for the feedforward dependency parser

[krickert]

The feedforward dependency parser from #1323 can now be trained, run, and evaluated from the command line, like the existing maxent DependencyParser tools.

Fixes #1469.

The new tools:

• FeedforwardDependencyParserTrainer trains a model from the same sample formats as DependencyParserTrainerME. The -params file takes the FeedforwardDependencyTrainer keys, and keys not given take the trainer's defaults. -refineParams refines the trained model with beam search, and -baseModel refines an existing model instead of training a new one.
• FeedforwardDependencyParser parses token_TAG lines from standard input and prints CoNLL-U.
• FeedforwardDependencyParserEvaluator reports UAS and LAS in the same format as DependencyParserEvaluator.

The parser and the evaluator take an optional -beamSize. Without it, they use the beam size a refined model records, or parse greedily.

opennlp FeedforwardDependencyParserTrainer.conllu -lang eng -data train.conllu \
    -model en-depparse-ff.bin -params ff.params -refineParams refine.params
opennlp FeedforwardDependencyParserEvaluator.conllu -model en-depparse-ff.bin -data dev.conllu
opennlp FeedforwardDependencyParser -beamSize 4 en-depparse-ff.bin < tagged-sentences.txt

ModelLoader now reports a model of another component as "Model has invalid format" for all CLI loaders, instead of failing with an IllegalArgumentException. The maxent and feedforward tools share the CoNLL-U output, scoring, and parameter loading code, and the maxent tools behave as before. There is no cross-validation tool, since no feedforward cross validator exists.

Tests cover the output of each tool, refinement, the beam size rules, and the rejected arguments, including exit codes, messages, and that no model file is written on failure. The dependency parser chapter of the manual gets a command-line section.

[krickert]

Fixing this now

[krickert]

rebasing