libc: malloc, calloc, realloc and memalign throw through the guest call on failure instead of returning NULL with ENOMEM; fix branch ready #5242

When an allocation fails, the guest malloc, calloc, realloc, memalign, aligned_alloc and reallocalign throw a C++ exception (bad_alloc, length_error, runtime_error or system_error) through the guest call, which ends the process because guest code cannot catch it. FreeBSD 11.4 (jemalloc) returns NULL with errno = ENOMEM, or EINVAL for an alignment that is zero or not a power of two. A guest that checks for NULL (the normal pattern) never gets the chance. Cases that abort today: a size the heap rejects (SIZE_MAX, PTRDIFF_MAX + 1, 1 << 40 and up), a calloc product that overflows, a bad alignment, an exhausted arena, and a size that fits the roughly 1 TB guest arena but exceeds the host's commit limit (malloc(300 GiB) throws "paging file too small").

Fix branch: https://github.com/husseinkad/AnyPS5/tree/fix/libc-alloc-failures (four commits on a recent main: the change, then three follow-ups from review). FileAndHeap.cpp catches the failures and sets errno; GuestHeap.cpp turns an arena runtime_error and the host out-of-memory codes (Windows 8, 14 and 1455, Linux ENOMEM) into bad_alloc, and releases the reserved arena range when a commit fails; GuestArena.cpp is left untouched on purpose so it does not conflict with #5218. New tests GuestAllocatorFailures.cpp and rewritten expectations in GuestHeapRounding.cpp and GuestReallocalign.cpp; debt entries in docs/dev/debt/libc.md.

Checks:

• guest_allocator_failures crashes on main and passes with the change; guest_heap_rounding, application_heap, guest_reallocalign, mspace, guest_memory and guest_new_handler pass. The test's leak check was weak in the first version (a reviewer removed the release call and it still passed); it now fails with that mutation.
• Probed through the guest exports against main: 1,000 failing 300 GiB mallocs all return NULL/ENOMEM with no throw, private bytes flat and the arena's lowest-free address unchanged; a failed realloc(p, huge) leaves p intact and malloc_usable_size unchanged, reallocf frees it and sets ENOMEM; alignments 0, 3, 24 and SIZE_MAX give EINVAL; errno is untouched on success over 1,000 mixed multi-threaded calls; with the arena filled to 1 MiB free every failing call gave ENOMEM; 6 threads allocating alongside 300 failing huge mallocs showed no corruption or deadlock; zero bytes stay committed after a failed commit. Side effect: posix_memalign and reallocf on arena exhaustion now return ENOMEM where main threw.
• Three independent reviews. check_conventions.py is clean; pr_overlap.py had no overlap on the first version, and fix(libkernel): fall back below a mapping hint with no room above it #5218 (which rewrites a function in GuestArena.cpp) is why that file is untouched.

Not covered: memalign and reallocalign setting errno is an assumption (FreeBSD has neither); FreeBSD 11.4 resizes realloc(ptr, 0) to a one-byte block, but this repo frees the block and returns NULL and existing tests assert that (recorded in the debt entry). The Linux commit-failure mapping was read, not built or run. Only Windows was built and run.

AI-assisted: yes