Add worstcase benchmark

[hendrikvanantwerpen]

Adds a benchmark that brings out the worst in tiktoken.

I tried to understand how the splitting regexes actually work, but that was kind of annoying.
Just selecting ranges from the whole Unicode set excluding whitespace seems to do the trick though.

If we want to ensure it really is the worst case, I can try again to understand the regexes, or we just reject inputs that can be split using the regex (assuming most cannot).

I've also added the graph for this benchmark to the readme.