Generators should use smaller discriminants

[jonas-schievink]

Currently they use a hardcoded u32 discriminant, which is quite a lot of room considering that generators don't typically have billions of states. They should be able to use the smallest unsigned integer type that fits all states instead (so u8 in the vast majority of cases).

[jonas-schievink]

The discriminant is also never marked with a valid range, meaning that it never provides any niches that could be used to store other values. Not sure if doing that has any drawbacks, but it seems like a somewhat easy size optimization on first glance.

[jonas-schievink]

Fixing this was way easier than I expected, but I'd appreciate help with benchmarking the runtime impact of the fix in #69837.

So if anyone knows any async/await benchmarks that are not I/O heavy (so that they mostly test CPU overhead of async/await), I'd be happy to run them.