Initial C-style interface for whisper.cpp

[ggerganov]

ref #9

First iteration of the C-style interface.
Already, a lot of the duplicated code in main.cpp and stream.cpp is now gone.

This is a high-level API that allows separate encode/decode calls + token sampling.
The user will be able to implement their own Mel computation, as well as their own token sampling algorithms using the computed token logits and probabilities.

For convenience, the greedy sampling (and in the future - the beam search sampling) are also available.

Additionally, there will be convenience functions that perform the entire interference (e.g. whisper_full) - simply provide RAW PCM audio and get the result.