Make model blocks compatible with any frame-wise or chunk-wise model

[juanmc2005]

Problem

The current implementation of FrameWiseModel and ChunkWiseModel is highly dependent on pyannote.audio internals and published models. Any model accepted by the pipeline must be loaded with pyannote.audio. This is a problem because users may want to plug in custom models that don't share the same structure as models trained with or published by pyannote.

Idea

Make FrameWiseModel and ChunkWiseModel accept a broad definition of a model, e.g. a Callable[[torch.Tensor], torch.Tensor] or something similar. This way, any model that can ingest a tensor (in a certain format) and return another tensor (also in a certain format) could be plugged in.

[juanmc2005]

Implemented in #61