* Hold on to the CUDA buffer - second try
see: https://github.com/davisking/dlib/pull/1855#discussion_r314666348

* Fix #1849 by calling device_global_buffer() unconditionally

* Added aarch64 to list of 64-bit architechtures

* Const-corrected declaration of ssyevr

When dlib is compiling, cmake will compare python architecture and target
architecture. So in cross-compiling case, it is irrevelant because host and
target architecture often differs. The main problem come from checking python
architecture on host and not on target.

Here is an error when compiling dlib from x86_64 to arm 32-bit target :
```
Python config failure: Python is 64-bit, chosen compiler is 32-bit
```

So :
- Skipping the comparation when cross-compiling is enabled.
Signed-off-by: Romain Naour <romain.naour@smile.fr>
Signed-off-by: Alexandre PAYEN <alexandre.payen@smile.fr>

Push all include and link options needed for dlib to pkg-config.  We do this by getting them from the same list cmake uses.

* Add point assignment test

* Fix setting points y coordinate changes x instead (issue #1794)

Apparently pytest is still sort of busted.

This reverts commit 5e63d015.

Add input_grayscale_image_pyramid

libjpeg not being able to open some types of jpeg file.

* Upgrades dlib's included libjpeg to version 8d

* Overloads load_jpeg to read from memory buffer

* Removes "__inline__" define in jconfig, broke VC build

* Changes buffer size type to size_t

These warnings occurred when building the semantic segmentation
examples

* prevent GCC from complaining about this unused parameter

* fix some pedantic warnings

* remove unneeded assert

given an empty list of initial function evaluations.

* Semantic-segmentation loss calculation: fix buffer usage on multi-GPU training

* Review fix: make the work buffer live longer