PS-10070, PS-10083 [8.4]: fixes for Apple clang

[jankowsk]

https://perconadev.atlassian.net/browse/PS-10070

rapidjson's TypeHelper has specializations for the fixed-width integer typedefs (uint64_t etc.) but not for size_t. On macOS / Apple clang + libc++, size_t is 'unsigned long' which has no TypeHelper specialization, so get_element<size_t>() fails to compile. Read into a uint64_t and assign back on success; uint64_t is recognized on all platforms.

[jankowsk]

Added another commit fixing problems with https://perconadev.atlassian.net/browse/PS-10083

The static_assert(sizeof(thread_group_t) == 512, ...) in sql/threadpool_unix.cc has existed since PS 8.0.12 and was hand-tuned against the Linux/glibc layout of the embedded mysql_mutex_t and PSI structures. PS-10083 ("Add more statistics for threadpool", merged 2026-03-11 between 9.6.0 and 9.7.0) inserted two LiveStats members into thread_group_t and adjusted the trailing 'char padding[]' from 328 to 248 to keep the Linux total at exactly 512.

LiveStats itself is platform-neutral (5 PODs, 40 bytes), but mysql_mutex_t expands by ~128 bytes on macOS / Apple clang + libc++, so on that platform the struct lands at 640 bytes and breaks the build:

sql/threadpool_unix.cc:155:1: error: static assertion failed due
to requirement 'sizeof(thread_group_t) == 512'
The false-sharing guarantee comes from alignas(128); what matters is that the struct occupies a whole number of 128-byte cache lines, not that it is exactly 512 bytes (640 is also a multiple of 128). Replace the equality check with 'sizeof % 128 == 0' so Linux/glibc, libc++ on macOS, and any future libstdc++ layout drift all pass while the no-false-sharing invariant remains enforced.