Remove Python 3.5 tests on AppVeyor because something is wrong with python 3.5 on appveyor and the other tests are sufficient for python coverage.

Apply the #1514 fix even on non-Windows platforms, and change all remaining cudaStreamSynchronize calls (#1596)

* Apply the #1514 fix even on non-Windows platforms

* Try to fix #1513 even more by circumventing the remaining cudaStreamSynchronize calls

* Make the fix apply only if CUDA_VERSION == 10000

* Make the fix apply on CUDA 9.2 also

* CHECK_CUDA(cudaStreamSynchronize(stream));

ECX must be set to 0 to read this flag, the easiest way to achieve this
is to use __cpuid_count with 0 as the second argument.

* Add concat_prev layer, and U-net example for semantic segmentation

* Allow to supply mini-batch size as command-line parameter

* Decrease default mini-batch size from 30 to 24

* Resize t1, if needed

* Use DenseNet-style blocks instead of residual learning

* Increase default mini-batch size to 50

* Increase default mini-batch size from 50 to 60

* Resize even during the backward step, if needed

* Use resize_bilinear_gradient for the backward step

* Fix function call ambiguity problem

* Clear destination before adding gradient

* Works OK-ish

* Add more U-tags

* Tweak default mini-batch size

* Define a simpler network when using Microsoft Visual C++ compiler; clean up the DenseNet stuff (leaving it for a later PR)

* Decrease default mini-batch size from 24 to 23

* Define separate dnn filename for MSVC++ and not

* Add documentation for the resize_to_prev layer; move the implementation so that it comes after mult_prev

* Fix previous typo

* Minor formatting changes

* Reverse the ordering of levels

* Increase the learning-rate stopping criterion back to 1e-4 (was 1e-8)

* Use more U-tags even on Windows

* Minor formatting

* Latest MSVC 2017 builds fast, so there's no need to limit the depth any longer

* Tweak default mini-batch size again

* Even though latest MSVC can now build the extra layers, it does not mean we should add them!

* Fix naming

the register storage specifier was deprecated in C++11 and removed in C++17

* Add pointwise_divide operator

* Add tests for pointwise_divide function

* Replace in affine layer pointwise_multiply for division by its equivalent pointwise_divide.

* Add DLIB_USE_MKL_WITH_TBB opcion to CMake for Windows.

* Add to CMake MKL with TBB support for Linux

* Remove unnecessary tbb includes from CMake.

* Add back white spaces.

copies of libpng or libjpeg, instead, use the copies what come with dlib.  We
are doing this because too many Visual Studio users have busted copies of these
libraries on their machines.  This will make things just always work.

* Problem: see #1513 (https://github.com/davisking/dlib/issues/1513)
Candidate solution: busy-loop until cudaStreamQuery returns cudaSuccess

* Make the suggested fix a Windows-only thing

* Fix for "warning C4724: potential mod by 0" in matrix.h

* Fixed random_forest_regression.h warnings

* Added weighted labels to loss_binary_log

* Added weighted labels to loss_binary_log

* Clarified docs.

* clarified docs

* added scale_image python binding which returns an image resized by the provided scale factor

* added scale_rect for rectangle and drectangle. returns a rectangle scaled by the provided scaling factor

* - do not cast scale_rect coordinates to long for drectangle
- document scale_rect for drectangle and rectangle in _abstract.h
- change python binding scale_image to resize_image
- update documentation for resize_image to describe behaviour

Add the __virtual_inheritance keyword to dummy struct when compiling with MSVC.