AVX v256_cleanup() fixed

[savuor]

This pullrequest changes

_mm256_zeroupper is replaced by _mm256_zeroall

This should fix an incorrect behavior of vx_cleanup on AVX+ baselines on some compilers (for example GCC 7.3.0, Ubuntu 18.04).

Details

_mm256_zeroupper unpredictably spoils local variables even when there's no other intrinsics used in the code.
_mm256_zeroall doesn't cause such errors at the same time having similar purpose and performance.

[alalek]

https://software.intel.com/en-us/articles/avoiding-avx-sse-transition-penalties

1. Problem is observed with AVX2 baseline only.
2. There are many of these vx_cleanup() calls in the middle of functions. This doesn't look correct, make no sense and not working.
3. Also "transitions" should be handled before AVX code too: SSE->AVX switch.

I believe we should follow these rules instead of blind hacking:

1. Any "exported" OpenCV function should make transitions calls before and after executing any code (CPU_BASELINE is AVX+). Ideal place for that is CV_INSTRUMENT_REGION(); macro.
2. AVX+ "dispatched" OpenCV functions should make transitions calls before and after executing any code (CPU_BASELINE is SSE*). Partially handled by __CV_AVX_GUARD macro (via CV_INSTRUMENT_REGION() - but it is missing in new updates). Perhaps this should be moved into CV_CPU_CALL_*() macros for dispatched calls.
3. Inline functions/operators should not call this at all.

[vpisarev]

@alalek, vx_cleanup() should be much cheaper (maybe even much cheaper) than CV_INSTRUMENT_REGION() call. So, I suggest to use vx_cleanup() explicitly where it's needed. Of course, CV_INSTRUMENT_REGION() may be changed as well, to play on the safe side. 👍