fix: free(NULL) crashes before initialization with the 2-level page map

[LucFauvel]

fixes anthropics/claude-code#89370

Issue not reachable on upstream microsoft/mimalloc as _mi_aligned_ptr_page0 early NULL returns, while this fork is always via page map.

The static mi_page_map_empty used before _mi_page_map_init runs has submaps[0] == NULL. The default (non-secure, non-debug) build resolves a pointer through _mi_unchecked_ptr_page, which loads submaps[idx][sub_idx] without testing the sub-map for NULL, so free(NULL) -- where both indices are 0 -- dereferences address 0 instead of yielding a NULL page.

_mi_checked_ptr_page does test for it, which is why only the default release build is affected; MI_SECURE / MI_DEBUG builds are fine. The flat page map got this right in microsoft#1341 by pointing its empty map at a real one-element array; this does the same for the 2-level map.

Reachable in practice: glibc >= 2.44 ends newlocale() with an unconditional free() of a buffer that is NULL on the common path, and the C++ runtime calls newlocale() from a static initializer while building std::locale::classic(). With MI_OVERRIDE that free(NULL) reaches mimalloc before any allocation has happened, so before the page map exists, and the process dies in a static initializer before main.

Codegen is unchanged (mi_free is byte-identical); the fix costs 8 bytes of BSS.

[LucFauvel]

Can't request review so @Jarred-Sumner if you can take a look. Seems to be breaking for more and more people on glibc 2.44. Upstream has made this bug unreachable (but latent) so an upstream sync might be better.

[LucFauvel]

Duplicate to changes in #30

Closing.