Allow LanguageDetector to be supplied with custom TestDataLanguageModel providers

[killergerbah]

Currently, the library loads its own language model from disk into memory. However, I would like to customize this behavior. For language detection to work, it looks like the only behavior required is an implementation of getRelativeFrequency for various n-gram language models.

This would allow me to have the option of storing the language model remotely (for example: in Redis), instead of needing hundreds of megabytes of memory for the in-memory language models.