for some reason Calib3d_Homography.fromImages on ARM gives much smaller number of inliers than on x86/x64. Need to investigate more carefully, but let's decrease the tolerance threshold for now

hack solvePoly to finds roots of polynoms with zero higher-order coefficients. The roots are populated in this case, which is not valid, strictly speaking, but good enough for function like correctMatches. This solves http://code.opencv.org/issues/4330

added tests for http://code.opencv.org/issues/4011 and http://code.opencv.org/issues/3057; fixed random subset generation in both methods to increase chance for a good subset

removed cameraMatrix modification in the currently disabled uPnP algorithm for SolvePnP problem (http://code.opencv.org/issues/3985)

added another triangulation test case from http://code.opencv.org/issues/3461; fixed partial derivative on distortion coefficients, as http://code.opencv.org/issues/4101 suggests

another attempt to make the MSER test pass. removed possible randomization in parameters from run to run

added reasonable test for MSER (including coverage for http://code.opencv.org/issues/4273, http://code.opencv.org/issues/1723 and http://code.opencv.org/issues/756); also, added some "in-progress" info to the Features2d_Feature2d.no_crash test.

Revert of "Merge pull request #4006 from sgjava:opencv-steve" (reverted from commit 47431840)

Revert of "Merge pull request #4014 from sgjava:opencv-sgjava" (reverted from commit 52fa55a7)