Extend r.watershed test suite for differences between ram and seg versions

[mwort]

These tests are designed to give a comprehensive overview of matching output between the in-memory (ram) and segmentation library versions of r.watershed with different optional inputs. It relies on running the module with both versions and a random raster (n=10000) for the optional inputs. The min/max of the difference between output rasters are compared and tests pass if they do not exceed the following precisions/thresholds (open for suggestions):

   output_precision = {
       'accumulation': 1,
       'tci': 0.01,
       'spi': 0.01,
       'drainage': 0,
       'basin': 0,
       'stream': 0,
       'half_basin': 0,
       'length_slope': 0.01,
       'slope_steepness': 0.01,
   }

The current state of matches with optional input in columns:

Output	required	flags=s	flags=4	depression	flow	disturbed_land	blocking	retention
accumulation	✅	✅	✅	🔴	✅	✅	✅	✅
tci	✅	✅	✅	✅	✅	✅	✅	✅
spi	✅	✅	✅	✅	✅	✅	✅	✅
drainage	✅	🔴	✅	✅	✅	✅	🔴	✅
basin	✅	✅	🔴	✅	🔴	✅	🔴	✅
stream	✅	✅	🔴	✅	🔴	✅	🔴	✅
half_basin	✅	✅	🔴	✅	🔴	✅	🔴	✅
length_slope	🔴	✅	🔴	🔴	🔴	🔴	🔴	🔴
slope_steepness	🔴	✅	🔴	🔴	🔴	🔴	🔴	🔴

I was not able to reproduce the mismatch in accumulation in the NC demo location reported by @HuidaeCho in #2222 with randomly assigned flow fractions (0-1), although other output does not match.

[HuidaeCho]

I think this test looks good. The only thing I would change is the number of reported failures. It's currently 1 from AssertionError if any tests fail. It would be useful to report the number of individual failed tests (red circles) somewhere.

[wenzeslaus]

I like the test and the output, however the main issue is that the test fails and we can automatically run only tests which pass when everything is all right. Maybe there is a number of expected failures here to test against?