Ensure CodePointTrie access by u16 or char omits unnecessary range checks

[hsivonen]

When CodePointTrie of the fast type is queried with u16, the code point cannot be above FAST_TYPE_FAST_INDEXING_MAX. When CodePointTrie (of either type) is queried with char, the code point cannot be above CODE_POINT_MAX.

We should ensure that the relevant branches are omitted in these cases. Perhaps its good enough to ensure this by Godbolting, but actually having separate code for these cases would be robust (not relying on the optimizer reaching a particular conclusion) and obvious.

[hsivonen]

Godbolting suggests that this optimizes the right way already. (Used trivial examples, not the actual CodePointTrie code.)

[sffc]

I think we should start with annotating these functions as #[inline]; ideally normalize_str is one single function that doesn't call down to other functions all the time. This should be sufficient for optimizing the branches even in -Os mode. If we need to use #[inline(always)] or a more invasive technique such as manually inlining or creating multiple versions of functions, we can do that if the benchmarks justify it.