Jesse Towner authored Feb 25, 2021 and Werner Lemberg committed Feb 27, 2021

This change fixes a crash that occurs in `Load_SBit_Png` when
running on a 64-bit Windows OS.  A memory access violation exception
would be raised by `setjmp` if the `jmp_buf` is not aligned to a
16-byte memory boundary.  This is due to setjmp executing `movdqa`
instructions to store 128-bit XMM registers to memory, which require
correct memory alignment.  This problem occurs because
`png_create_read_struct` uses `malloc` and `free` for memory
management, which only guarantees 8-byte alignment on Windows.

Instead, to fix the problem, `png_create_read_struct_2` is used on
64-bit Windows, which allows for user-defined memory allocation and
deallocation callbacks to be specified.  These callbacks forward the
allocation and deallocation requests to `_aligned_alloc` and
`_aligned_free`, ensuring that the allocated `png_struct` and
internal `jmp_buf` have the requisite 16-byte alignment.

* src/sfnt/pngshim.c <_WIN64>: Include `malloc.h`.
(malloc_callback, free_callback) <_WIN64>: New functions.
(Load_SBit_Png) <_WIN64>: Use `png_create_read_struct_2` instead of
`png_create_read_struct`