Current implementation of miniflann is releasing the trained index, and
rebuilding the index from the beginning.
But, some indexing algorithms like the LSH are able to add the indexing
data after that.
This branch is implementation of that optimization for LshIndex
FlannBasedMatcher in the feature2d.

(cherry picked from commit 8910508b)

* cmake CMP0054 and CMP0045
* aarch64 build
* portable code options in PCH cmake
* some of gcc 4.9 warnings

- updated existing Doxyfile.in
- added corresponding cmake instructions
- added some specific files (layout, icon)
- clean existing doxygen warnings

Conflicts:
	CMakeLists.txt
	doc/CMakeLists.txt
	modules/core/include/opencv2/core.hpp
	modules/core/include/opencv2/core/base.hpp
	modules/core/include/opencv2/core/cuda.inl.hpp
	modules/core/include/opencv2/core/mat.hpp
	modules/core/include/opencv2/core/matx.hpp
	modules/core/include/opencv2/core/types.hpp
	modules/flann/include/opencv2/flann/lsh_table.h
	modules/imgproc/include/opencv2/imgproc.hpp

Add the ensureSimpleDistance() method to ensure the user the returned distance is not ^2 (the default for L2 for instance)

Allows to choose orthogonal sub-vectors for LSH without using a static table among LshTable instances

version of print_params for backwards compatibility.

Fix: freeing previous elements has to be done before loading new parameters to avoid trying to delete unexisting objects if arrays size was modified

Avoid obtaining several identical dimensions between two LSH sub-vectors by choosing orthogonal sub-vectors.

When a cluster is empty for KMeans, it's better to give it the point from another cluster j that is the furthest one from center j.

Move templates in dist.h in order to share them between KMeansIndex and HierarchicalClusteringIndex classes.

As some processed distances are already ^2, use template to select whether or not we have to ^2 in KMeanspp

Pick centers in KMeans++ with a probability proportional to their distance^2, instead of simple distance, to previous centers

We don't use -Wcast-align ourselves, but the headers trigger warnings
for those users who do, which is not nice.

This is fixed upstream in mariusmuja/flann@b615f26, but that fix
would break binary compatibility, so I had to make a different one.

Since the bug isn't quite obvious, here's an explanation. In the const
version of any::cast, if policy is a small_any_policy, its get_value
returns its input argument. So r becomes a pointer to obj, and the
return value is a reference to a local variable, which is invalidated
when the function exits.

Looks like we don't actually use them; when creating precompiled headers
with Visual C++, a dummy .cpp is created at build time.

Fixed flann build with NEON;
Fixed Haming distance with NEON;
Honest cvRound for WinRT added;
cvRound test added;
Video IO with direct show disabled;

Invalid links and refences fixed.
SpaeseMat class documentation updated.

fixed possibly inccorrect interval measurements in flann's timer (bug #1165). Note: this may break binary compatibility, but since the class is not wrapped in Java and not exposed in Windows API, for example (it's considered internal-use class for flann), the effect should be minimal