Add websocket compatibility + other improvements

[juanmc2005]

This PR addresses issue #9, #73, #79 and partly #68.

Changelog

• Add WebSocketAudioSource
• Add websockets>=10.3 dependency
• Add websocket tutorial in the README
  • The tutorial works given Get an RTTM string from an Annotation object pyannote/pyannote-core#80 (pyannote.core>=4.5)
• Add dynamic resampling when sample rates don't match instead of failing
• Add Resample block
• Add --segmentation and --embedding arguments to diart.stream, diart.benchmark and diart.tune
  • These are only compatible with pyannote models. Loading models outside of pyannote are left for a future version
• Add diart.blocks.AdjustVolume to dynamically change the volume of chunks
• Make TemporalFeatureFormatter more flexible
  • It now allows the same instance to be used with different feature types
• Add attach_hooks() and attach_observers() to RealTimeInference so users can receive pipeline outputs and perform custom operations without rewriting the inference code
• Add to_dict() and to_inverse_dict() to SpeakerMap to handle assignments more easily
• Replace Optimizer.optimize(...) with Optimizer.__call__() to be consistent with the rest of the API
• Add possibility to set latency="max" or latency="min" in PipelineConfig
• Add possibility to set config=None in OnlineSpeakerDiarization, which will instantiate a default config
• Add diart.sinks.DiarizationPredictionAccumulator to accumulate diarization predictions in memory without writing to disk
• Make DiarizationPredictionAccumulator and RTTMWriter accept any tuple as long as the first element is an Annotation

[juanmc2005]

TODO: Rename RTTMAccumulator to something like DiarizationPredictionAccumulator.

This object has nothing to do with RTTM files

[juanmc2005]

Test the following:

• Websocket server and client
• Dynamic resampling with diart.stream
• --segmentation and --embedding
• AdjustVolume
• diart.benchmark
• diart.tune