Fix UB in number tokenizer: use pointer arithmetic instead of derefer…

[rayokota]

…encing end iterator

Summary

The number-literal scanner in Tokenizer.cpp passed &*byteStart and &*byteEnd to std::regex_search, where byteEnd is path_.cend(). Dereferencing a past-the-end iterator (*path_.cend()) is undefined behavior, even when only its address is taken.

Most standard libraries (libstdc++, libc++) happen to tolerate this, but MSVC's checked iterators (_ITERATOR_DEBUG_LEVEL) correctly reject it, causing an assertion failure / test crash on Windows MSVC 2022 builds.

Fix

Form the regex range from path_.data() pointers instead of dereferencing iterators:

const char* byteStart = path_.data() + byte_offsets_[position_];
const char* byteEnd   = path_.data() + path_.size();
std::cmatch match;
if (std::regex_search(byteStart, byteEnd, match, numregex) && ...)

data() + size() is a valid one-past-the-end pointer (legal to form, never dereferenced), which is exactly what std::regex_search expects for its [first, last) range. std::cmatch already operates on const char* ranges, so the surrounding match handling is unchanged. No behavioral change on conforming platforms — this only removes the UB that MSVC flags.

Testing

• All existing generated tests pass.

Reported by @u19809 while building with MSVC 2022 on Windows.