Fix performance degradation with -m32

[binhdvo]

Addresses performance issues caused by gcc failing to properly inline ZSTD_memmove when used inside ZSTD_copy16 macro. This change (from the original ZSTD_memcpy) was necessary because changes to the dctx handling now make overlap in memory a possibility. That original change was introduced in 6a7ede3 and had a negative impact of about 15% when compiling with the -m32 flag and using gcc.

Performance when applying this change to that commit is reduced to a 5% drop for -m32 using gcc. Performance improvement of this fix when applied to the current HEAD and using -m32 is 10%.

Performance testing indicates the loss was primarily due to this inlining behavior and not alignment issues. Continuing to try different approaches to see if there is a way to recoup the remaining performance loss.

[terrelln]

What does the x86-64 performance look like with this change?

[binhdvo]

What does the x86-64 performance look like with this change?

It was neutral as compared to the current HEAD

[ghost]

There is a SSE2 path:

#elif defined(ZSTD_ARCH_X86_SSE2)
    _mm_storeu_si128((__m128i*)dst, _mm_loadu_si128((const __m128i*)src));

But even use -m32 -msse2 options, gcc doesn't use that path:
https://godbolt.org/z/Yz3qsaca7

~~ Maybe ZSTD_ARCH_X86_SSE2 macro can be improved to support 32bit+SSE2.~~

edit: -m32 -msse2 shoudl enable ZSTD_ARCH_X86_SSE2 macro.
It seems clang use SSE2 by default, gcc doesn't.

[Cyan4973]

~~ Maybe ZSTD_ARCH_X86_SSE2 macro can be improved to support 32bit+SSE2.~~

-m32 -msse2 enables ZSTD_ARCH_X86_SSE2.
But -m32 alone doesn't (and shouldn't, since x86 cpus without sse2 exist, in contrast with x64).
Verified for both gcc and clang.