Cache engine generated from PFA documents

[ludovicc]

Parsing PFA is costly as it requires converting the document to an AST then compile Java code.

Suggested cache: https://github.com/ben-manes/caffeine

Issue: during cross-validation, many similar but different models are produced. They mostly differ only by some weights, but the underlying algorithm remains the same. The cache would be more effective if it could reuse the compiled algorithm from previous executions.