Fix segfault when built with -mavx (#108)

[kpchoi]

XEVE_CTX embeds XEVE_PINTRA/XEVE_PINTER, which contain ALIGNED_32 members,
so the required alignment of the type is 32 bytes. It is however allocated with
plain malloc(), which only guarantees 16 bytes. When built with -mavx, the
compiler relies on the declared alignment and merges adjacent member stores into
aligned vmovdqa instructions, which fault at runtime. This is what crashes
xeve_pintra_create() in #108.

This adds xeve_malloc_align32() / xeve_mfree_align32() and uses them for the
XEVE_CTX / XEVEM_CTX allocations. Note that no SIMD kernel actually requires
aligned buffers (all intrinsics use unaligned load/store), so no other allocation
is affected.

Verified with cmake -DCMAKE_C_FLAGS="-mavx": the crash reproduces before the
change and is gone after, for both MAIN and BASE profiles, single- and
multi-threaded. The produced bitstream is bit-exact with a non-AVX build.

Fixes #108