270996c introduces Multiprocessing after the Multithreading by Rust tokenizers.

Although Multithreading should be finished by then, forking processes in python results in:

huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks...
To disable this warning, you can either:
	- Avoid using `tokenizers` before the fork if possible
	- Explicitly set the environment variable TOKENIZERS_PARALLELISM=(true | false)

When setting the env var to false, python mp wont start.

My next experiment will be adding mp on a higher level (calling dataset_from_dicts) again.

I did some performance benchmarking and found the culprit:
The function offset_to_token_idx() in samples.py takes up 93% of compute time.
Vectorizing the function reduced preprocessing from 199 secs to 21 secs!

I will now work on separating the create_samples_qa and sample_to_features_qa functions into more meaningful functions:

1. split_documents_into_passages
2. featurize_text_and_labels

I tested this branch with haystack in the following ways:

• running all haystack tests
• Tutorial 1
• Tutorial 5, when using FarmReader.eval we need to adjust the return of processor.dataset_from_dicts
• Colab tutorial 5

Working on fixing the remaining tests.

I cannot reproduce the failing s3 test and nothing has changed code wise.

I presume it is some CI problem that we can fix later.

Merging now.