Add python API for diart.stream and diart.benchmark

[juanmc2005]

Problem

Having these scripts is nice but it requires a lot of work to swap pipeline components. For example, how do I benchmark a pipeline with an embedding model from an external library like speechbrain? This currently requires to implement the embedding model, build a new pipeline from scratch and then change diart.benchmark's code, which is not very practical.

Idea

It would be great to have a generic pipeline where components can be swapped easily. Getting there takes a lot of development though, so in the meantime let's at least provide a python API to run streaming and benchmarking out-of-the-box.

This should also facilitate #53.

Example

import torch
from diart.benchmark import Benchmark

gpu = torch.device("cuda:0")
pipeline = SomeCustomPipeline()

benchmark = Benchmark("/path/to/speech", "/path/to/rttm", "/path/to/output", batch_size=32, device=gpu)

benchmark(pipeline)

[juanmc2005]

Implemented in #55.
See docstrings for RealTimeInference and Benchmark for the final API.