Auto merge of #109216 - martingms:unicode-case-lut-shrink, r=Mark-Simulacrum

Shrink unicode case-mapping LUTs by 24k

I was looking into the binary bloat of a small program using `str::to_lowercase` and `str::to_uppercase`, and noticed that the lookup tables used for case mapping had a lot of zero-bytes in them. The reason for this is that since some characters map to up to three other characters when lower or uppercased, the LUTs store a `[char; 3]` for each character. However, the vast majority of cases only map to a single new character, in other words most of the entries are e.g. `(lowerc, [upperc, '\0', '\0'])`.
This PR introduces a new encoding scheme for these tables.

The changes reduces the size of my test binary by about 24K.

I've also done some `#[bench]`marks on unicode-heavy test data, and found that the performance of both `str::to_lowercase` and `str::to_uppercase` improves by up to 20%. These measurements are obviously very dependent on the character distribution of the data.

Someone else will have to decide whether this more complex scheme is worth it or not, I was just goofing around a bit and here's what came out of it 🤷‍♂️ No hard feelings if this isn't wanted!

Author: bors