[FIX] relief: Fix contingency (de)allocation

[ales-erjavec]

Issue

_contingency_table in relief.pyx does not retain np.ndarray allocated data buffer after function returns even though it 'inserts' a cython MemoryView into a c++ std::map, which is then used later (I think MemoryView cannot be readily inserted in c++ std containers and preserve referential integrity).

This can lead to a segfault or memory corruption. It can be triggered on macOS with

MallocScribble=1 PYTHONMALLOC=debug python -X dev -m unittest -v -b  Orange.widgets.data.tests.test_owrank.TestOWRank.test_dataset{,,,,}

Description of changes

Use Dict[int, np.ndarray] instead of std::map[ssize_t, MemoryView] as 'contingencies'. Acquire GIL when necessary when accessing elements.

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #4745 into master will increase coverage by 0.05%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #4745      +/-   ##
==========================================
+ Coverage   83.69%   83.74%   +0.05%     
==========================================
  Files         275      275              
  Lines       55620    55655      +35     
==========================================
+ Hits        46551    46611      +60     
+ Misses       9069     9044      -25