cxxpiper improvement

https://github.com/brailcom/speechd/pull/996 introduced a cxxpiper module, but it needs improvement, as mentioned in the comments, but which have been unanswered, so copy/pasting them here so they are not lost:

- it will be nice long-term to be able to load several models. Even better, list all of them but load them only on-demand, so people don't have to modify any file but just install a model package and see the language pop-up in orca
- It's quite sad that piper doesn't provide parsePhonemizeConfig, parseSynthesisConfig, parseModelConfig, initialize, loadModel, loadVoice, synthesize, textToAudio, as library functions. It'd be good to tell upstream to do so, rather than copying their code
- we'd want to see piper support pipelining, to start sending audio to the server before the whole audio is produced. That would be really important to get good reactivity
- It would be useful to make CUDA support work