Use memcpy for unaligned reads to appease UBSan

[SpexGuy]

This change uses memcpy to perform unaligned reads, instead of reinterpret_cast. In tests on godbolt, this generated a single unaligned read on x86 and x86_64 platforms even under -O0. Other platforms generate byte reads and then stitch them together with shifting. Writing it this way removes the need for the build to force byte reads when UBSan is enabled. However, unfortunately FPNG_USE_UNALIGNED_LOADS is still needed for performance in READ_RGB_PIXEL, because most compilers cannot remove the vestigial fourth byte read on platforms that require aligned reads.

[SpexGuy]

Hmm, actually it looks like MSVC takes a little more convincing, it will actually call memcpy on /O1, it needs /O2 to generate good code. Depending on how important debug mode performance is on MSVC, this may or may not be worth integrating.

[richgel999]

Thanks - let me look at this and profile it. (This stuff is way trickier than it should be in C/C++.) Not concerned very much with debug perf, as long as it's good enough.