Add speed benchmarking pipeline

[tanaysoni]

This PR adds an automated speed benchmarking pipeline using pytest-benchmark.

As a first benchmark, a test case is added comparing inference latency for Question Answering using PyTorch vs ONNX against different combinations of batch_size and max_seq_len.

Additionally, this PR also fixes the test for multiprocessing in the Inferencer.

TODO

• writing output to json (and potentially in a spreadsheet)
• dockerfile for onnx-gpu
• add param for length of doc (140k=40 pages, 35k=10 pages, 3.5k => 1 page)
• modularizing the benchmark further so that you can run pytorch and onnx, cpu and gpu separately

[tholor]

Would be great to add:

• a small readme with instructions on how to run (e.g. which env needed for onnx?)
• writing the results somewhere to a file / table than we can link in the repo and update after releases / major changes

[tholor]

Looking good!