Add benchmark vs tiktoken

[Narsil]

Before making all the optimizations it's better to start with a clear picture of where we stand.

Benchmarking code checks llama-3.1 (which uses a regex splitting macro) against XNLI in 3 scenarios.

Single sentence (inference use case very often).
Many sentences (dataset).
Many sentences fused into a large blob (advanced use case potentially showcasing different profile).

==============
num_threads: 1, data size: 24.59 KB, documents: 10
tiktoken 	4.25 MB  / s
huggingface 	1.93 MB / s
==============
num_threads: 1, data size: 24.04 MB, documents: 10000
tiktoken 	5.17 MB  / s
huggingface 	2.05 MB / s
==============
num_threads: 1, data size: 24.04 MB, documents: 1
tiktoken 	5.88 MB  / s
huggingface 	1.28 MB / s
==============
num_threads: 2, data size: 24.59 KB, documents: 10
tiktoken 	6.49 MB  / s
huggingface 	1.88 MB / s
==============
num_threads: 2, data size: 24.04 MB, documents: 10000
tiktoken 	9.69 MB  / s
huggingface 	2.12 MB / s
==============
num_threads: 2, data size: 24.04 MB, documents: 1
tiktoken 	5.90 MB  / s
huggingface 	1.35 MB / s
==============
num_threads: 4, data size: 24.59 KB, documents: 10
tiktoken 	9.26 MB  / s
huggingface 	1.88 MB / s
==============
num_threads: 4, data size: 24.04 MB, documents: 10000
tiktoken 	14.71 MB  / s
huggingface 	2.16 MB / s
==============
num_threads: 4, data size: 24.04 MB, documents: 1
tiktoken 	5.67 MB  / s
huggingface 	1.39 MB / s
==============
num_threads: 8, data size: 24.59 KB, documents: 10
tiktoken 	7.35 MB  / s
huggingface 	1.95 MB / s
==============
num_threads: 8, data size: 24.04 MB, documents: 10000
tiktoken 	18.52 MB  / s
huggingface 	2.16 MB / s
==============
num_threads: 8, data size: 24.04 MB, documents: 1
tiktoken 	5.73 MB  / s
huggingface 	1.36 MB / s

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.