Use Kaldi GrammarFST on the fly decoding

[mathquis]

To avoid the language model being biased by unbalanced intents with multiple slots and help reduce Kaldi decoder resource usage, the implementation of Kaldi's GrammarFST approach seems like a good solution.
https://kaldi-asr.org/doc/grammar.html

Slots should be compiled as separate wFST and the intents are compiled as another wFST. Everything is stitched up at runtime by the Kaldi decoder so the resource usage is reduced (only the intents graph is loaded, the slots graphs are loaded on the fly).

Using this means that the weight of play music by (ARTIST) will be the same for any value in the artists slot and should greatly help avoiding incorrect graph paths due to 3-grams model issues.
See: synesthesiam/rhasspy#82