Resize reworked using wide universal intrinsics (#13781)

* Added wide universal intrinsics optimized implementation for 3 channel bit-exact linear resize

* Reworked linear resize using new wide LUT intrinsics

* Fix for VSX intrinsics

All <arm_neon.h> includes in core/cv_cpu_dispatch.h are protected by an
ifndef __CUDACC__ to prevent attempting to use neon intrinsics when
compiling cuda kernels (.cu) -- this prevents hard errors such as
  error: identifier "__builtin_neon_qi" is undefined

Add this same protection to flann/dist.h to fix compilation involving
flann.hpp.

Fix arguments parsing. Add possibility to choose between different features type. Add keypoints matching visualization. Auto format code.

- gemm: keep baseline only (lapack is 10x+ faster, lets reduce binary size)
- transform / distTransform
- scaleAdd (32f/64f only)
- Mahalanobis: keep baseline only (no perf tests)
- mulTransposed: keep baseline only (no perf tests)
- dot

- Fix hanging issue on 2080 Ti
- Fix racing issue

details: https://github.com/opencv/opencv/pull/13726

In order to help CMake and PkgConfig find FFmpeg you need to:
1) export `PKG_CONFIG_LIBDIR=${SYSROOT}`
where `SYSROOT=${ANDROID_NDK}/sysroot`
2) Add `-DCMAKE_PREFIX_PATH=${FFMPEG_INSTALL_PREFIX}`
3) Append `-DCMAKE_FIND_ROOT_PATH=${FFMPEG_INSTALL_PREFIX}`

* re-enabled NEON routines for arm64