Merge vocabulary and context building

[bunny-therapist]

Partially addresses #46

Context and vocabulary building each had their own separate loop over all tokens, and tag functions were called multiple times for the same word. Occurrences stored shifts and offsets even though the only thing that was relevant was whether it was at the start of a sentence or not.

This merges context and vocabulary building into one function, one loop, that calculates the tag for each occurrence only once.

I suspect that extract_features and filtering could also be merged into the same loop.

I assume that reducing the number of loops and function calls is a performance improvement, but I will rely on @xamgore to verify this :)

[bunny-therapist]

We can easily make "tags to discard" configurable and just use that here. LIAAD/yake wanted that, but never got to it.

[bunny-therapist]

I just noticed that the CI checks are duplicated. There is one for pull_request and one for pull_request target. Completely unnecessary. I will look into how to fix it.

[xamgore]

I like it! Allocations are a better as well as the speed. 4-6%

before

bench          fastest       │ slowest       │ median        │ mean          │ samples │ iters
├─ text_3kb    3.523 ms      │ 4.558 ms      │ 3.572 ms      │ 3.579 ms      │ 27925   │ 27925
├─ text_100kb  122.1 ms      │ 130.2 ms      │ 124.2 ms      │ 124.6 ms      │ 803     │ 803
╰─ text_170b   159.2 µs      │ 391 µs        │ 162.7 µs      │ 164 µs        │ 609118  │ 609118

after

bench          fastest       │ slowest       │ median        │ mean          │ samples │ iters
├─ text_3kb    3.378 ms      │ 22.48 ms      │ 3.416 ms      │ 3.424 ms      │ 29191   │ 29191
├─ text_100kb  115.7 ms      │ 120.3 ms      │ 116.5 ms      │ 116.5 ms      │ 858     │ 858
╰─ text_170b   151 µs        │ 357.6 µs      │ 153.5 µs      │ 154.1 µs      │ 648169  │ 648169

[xamgore]

With my latest changes looks like we won another 10% with this PR. On my laptop then 1kb = 1ms. If I plug power in, even better.

bench          fastest       │ slowest       │ median        │ mean          │ samples │ iters
├─ text_3kb    3.122 ms      │ 4.182 ms      │ 3.173 ms      │ 3.193 ms      │ 18775   │ 18775
├─ text_100kb  104.4 ms      │ 109.4 ms      │ 105.4 ms      │ 105.8 ms      │ 567     │ 567
╰─ text_170b   144 µs        │ 305.7 µs      │ 146.8 µs      │ 148.6 µs      │ 67226   │ 67226