Expand Python API capabilities

[eginhard]

This PR aligns the Python API more closely with what is available via the CLI. Previously pretrained TTS models could only be used with their default vocoder.

For example, this uses vocoder_models/en/ljspeech/hifigan_v2

from TTS.api import TTS

tts = TTS("tts_models/en/ljspeech/fast_pitch")
_ = tts.tts("hello")

Now you can also pass a different pretrained vocoder name (fixes coqui-ai#3558):

tts = TTS(
    "tts_models/en/ljspeech/fast_pitch",
    vocoder_name="vocoder_models/en/ljspeech/multiband-melgan"
)
_ = tts.tts("hello")

Combining a pretrained TTS model with a local vocoder and vice-versa or passing a custom speaker encoder is also possible now.

I then refactored the CLI to use the Python API internally, so that everything goes through the same pipeline for consistency.