Dynamic switch at runtime between SSE2 and AVX2 optim of IDWT 5x3

[rouault]

#957 has a function that can be compiled either with SSE2 or AVX2. It could be desirable to have binaries where the function is compiled in the 2 versions, and depending on the availability of the instruction set, decide which version to use.

Caution: do not forget to add a call to _mm256_zeroupper() after the last use of the AVX/AVX2 instruction set, to avoid transition penalties when returning to SSE code. See https://software.intel.com/en-us/articles/avoiding-avx-sse-transition-penalties
(This is not needed when compiling the code entirely with AVX2 enabled, since SSE code used the VEX mnemonics, which avoid this)