Online speaker diarization as a block

[juanmc2005]

This PR addresses issues #83 and #84.

Changelog

• OnlineSpeakerDiarization is now independent from RxPY and can be used as a block (cc @hbredin)
  • It can receive a single waveform or a list of them (for batched inference)
  • This class is now stateful (speaker centroids), so it now allows to initialize the system with centroids from a previous pipeline
• from diart import OnlineSpeakerDiarization, PipelineConfig
• AudioLoader cannot split files into audio chunks anymore
• Emit additional chunks with zero padding at the end of file streams so the pipeline output is fully aggregated
  • Before, this was done by notifying DelayedAggregation of the stream duration, then it would concatenate the last non-aggregated output
• Fix bug: EmbeddingNormalization was squeezing the output
• Uris are no longer required by blocks or output annotations, instead they're added by sinks
• OnlineSpeakerDiarization does not split the stream into chunks anymore, nor it resamples chunks dynamically
  • This is now done by RealTimeInference, which still uses RxPY as it is a higher level API
• RealTimeInference can now do batched inference and includes new parameters
• Benchmark reuses RealTimeInference internally (huge win here)
• regularize_audio_stream() renamed to rearrange_audio_stream(), as the notion of regularity is not very clear here
• diart.pipelines does not exist anymore
• Audio sources don't have a length property anymore
• Remove PrecalculatedFeaturesAudioSource
• Customizable reading block size in most audio sources
• When possible, AudioSource block size is set to the step size
• Add OnlineSpeakerDiarization.reset() to reset internal state (centroids and buffers)
• Add AudioSource.close() to correctly handle termination from external causes
• Avoid confusion with Rx observers not being called with do and do_action
  • Replace diart.operators.profile with a stateful diart.utils.Chronometer
  • Absorb diart.operators.progress in RealTimeInference
  • RealTimeInference now handles all the complexity of Rx

RAM usage during inference is considerably reduced (~30%)
Runtime doesn't seem to be impacted