WARNING: THIS PR WILL BREAK USER CODE IN CASE THEY:

• UTILIZE HARD-CODED MULTI-CHANNELS NUMERIC VALUES INSTEAD OF PROPER TYPE VALUES (E.G.: 8 FOR CV_8UC2)
• UTILIZE HARD-CODED NUMERIC INCREMENTS TO STEP FROM A CHANNEL TO ANOTHER (E.G. CV_32FC2 + 8 for CV_32FC3)

Merge with opencv/opencv_contrib#1790
resolves #12584
related #12569

This pullrequest changes

• Increase the supported depth types limit from 8 to 32
  • Introduced uint32, uint64, int64, Q formats, and other special formats (i.e. raw, auto, undefined)
• Increase the supported channels limit from 512 to 2048 (still extensible with further design work)
  • Use the next available (bigger) number of channels in case of invalid channel input
  • Great impact expected on various modules (e.g. DNN and feature2D)
• Fully-backwards compatible with serialized files generated using older OpenCV versions

Future work:

• assertion for (cn)==(cn) in CV_MAKETYPE to avoid random input multiple evaluation
• replace expressions like depth = rng.uniform(CV_8U, CV_16F+1); with:

vector<int> depths {CV_8U, CV_16U, CV_32U, CV_64U, CV_8S, CV_16S, CV_32S, CV_64S, CV_32F, CV_64F};
depth = depths[rng.uniform(0, depths.size())];

• utilize and test new types
• internally support both CV_32F and CV_64F with old and new numeric values, while defaulting to the new values

force_builders=Custom,Linux32,Win32,Win64 MSVS2017,ARMv7,ARMv8,Android pack
docker_image:Docs=docs-js
docker_image:Custom=ubuntu-cuda:16.04