Motivation:

#17007 fixed #13749 by making
`AsciiString.cached(String)` sanitize non-Latin-1 input before caching
the `String` value.

That change preserved the correct behavior, but introduced a performance
regression. The implementation now performs two separate scans over the
input string:

1. `new AsciiString(string)` copies chars into the backing byte array
via `c2b(...)`.
2. `AsciiString.cached(String)` scans the same string again to check
whether any char is greater than `MAX_CHAR_VALUE`.

PrintAssembly results show that C2 does not merge these two loops after
JIT compilation. The old version still contains the second scan loop in
`AsciiString.cached`, while the optimized version has a single loop that
performs both `c2b(...)` conversion and the Latin-1 check.


[jit-actual-old-c2-ascii-final.txt](https://github.com/user-attachments/files/29848724/jit-actual-old-c2-ascii-final.txt)


[jit-actual-current-c2-ascii-final.txt](https://github.com/user-attachments/files/29848711/jit-actual-new-c2-ascii-final.txt)

Modification:

Merge the byte conversion and Latin-1 detection into a single loop in
`AsciiString.cached(String)`.

The method now allocates the byte array directly, fills it using the
existing `c2b(char)` conversion, and tracks whether every input char is
Latin-1 during the same pass. It still preserves the original `String`
instance for ASCII/Latin-1 input, and reconstructs the cached string
from the byte array only when sanitization is required.


Result:
This removes the extra input scan while preserving the behavior
introduced by #17007.