Optimize Reader.takeLeb128

Rewrite `Reader.takeLeb128` to not use `takeMultipleOf7Leb128` and
instead:
 * Use byte aligned integers
 * Turn the decoding into a finite state machine
 * Turn the main reading loop into an inlined loop of static length
 * Special case small integers (<= 7 bits)

Across larger integers (> 7 bits) this roughly doubles performance in a
[micro-benchmark](https://zigbin.io/b8b396), however with some notable
exceptions:
 * `i50` sees barely any speedup and `u50` even sees a minor slowdown. I
   have genuinely no clue what could be causing this, it's the exception
   to the rule.
 * 8, 16 and 32 see a significanly higher speedup

Also expand on test coverage

Author: NicoElbers