Now it allows choosing between arbitrary implementation variants.

now DeviceInfo provides full information about device (from cudaDeviceProp)

Use min/max bounds instead of local value to calculate acceptance threshold.
Threshold based on local values somethimes does not work because cancellation
of big values may produce error bigger than local value.

This option provides an easy way to identify tests having no regression data
for sanity checks

Also fixed a typo in performance testing framework and removed hardcoded
temporary file name from highgui perf test

This reverts commit 54a202b3.

Conflicts:
	modules/core/doc/command_line_parser.rst
	modules/core/include/opencv2/core/core.hpp
	modules/core/src/cmdparser.cpp
	modules/gpu/perf/main.cpp
	modules/gpu/test/main.cpp
	modules/ts/src/ts_perf.cpp