fix: use manual padding of instance_map_shard (#5200)

* Use manual padding of instance_map_shard.

The alignas(64) specifier requires aligned allocation, which is not
available on macOS when targeting versions before 10.14.

* Add 'see #5200'

* Update include/pybind11/detail/internals.h

* Update include/pybind11/detail/internals.h

---------

Co-authored-by: Henry Schreiner <HenrySchreinerIII@gmail.com>

Author: colesbury

include/pybind11/detail/internals.h

@@ -148,18 +148,14 @@ struct override_hash {
 
 using instance_map = std::unordered_multimap<const void *, instance *>;
 
-// ignore: structure was padded due to alignment specifier
-PYBIND11_WARNING_PUSH
-PYBIND11_WARNING_DISABLE_MSVC(4324)
-
 // Instance map shards are used to reduce mutex contention in free-threaded Python.
-struct alignas(64) instance_map_shard {
+struct instance_map_shard {
     std::mutex mutex;
     instance_map registered_instances;
+    // alignas(64) would be better, but causes compile errors in macOS before 10.14 (see #5200)
+    char padding[64 - (sizeof(std::mutex) + sizeof(instance_map)) % 64];
 };
 
-PYBIND11_WARNING_POP
-
 /// Internal data structure used to track registered instances and types.
 /// Whenever binary incompatible changes are made to this structure,
 /// `PYBIND11_INTERNALS_VERSION` must be incremented.