Benchmark script + improvements and bug fixes

[juanmc2005]

This PR addresses issues #35 and #39

Changelog

• Add diart.benchmark script for fast inference, evaluation and real-time latency estimation of the pipeline (issue Add batched mode for faster inference on pre-recorded conversations #35). Only for the standard pipeline for now, custom pipelines will have to modify some internals until this is improved.
• Add GPU compatibility for diart.benchmark and diart.demo with --gpu
• Add real-time latency estimation to FileAudioSource with the profile argument (issue Benchmarking operator #39)
• Add ChunkLoader to centralize all file chunking
• Add tqdm progress operator: observable.pipe(dops.progress()).subscribe(...)
• Add PipelineConfig to encapsulate the configuration of the pipeline
• Add --no-plot argument to diart.demo to skip plotting (as mentioned in Question about reproduce the result #36)
• Make blocks in functional compatible with batched inference
• Fix resolution bug in FrameWiseModel always using chunk duration from model training instead of the duration of the given chunk
• Make models compatible with torch.Tensor and numpy.ndarray inputs (SlidingWindowFeature compatibility is kept but only for non-batched mode)
• Add OnlineSpeakerTracking as a submodule of the pipeline including clustering and output reconstruction
• Add from_file() method to OnlineSpeakerDiarization to run batched inference, as opposed to from_source() that runs online inference
• Sources can now provide the number of chunks that they will emit (if known) via the length: Optional[int] property

Notes

• Benchmarked performance on AMI test with latency=5: DER = 27.3
• Benchmarked performance on VoxConverse test with latency=5: DER = 17.1