[ENH] Faster normalization

[markotoplak]

Issue

Fixes #5219

Description of changes

The computation of distributions is slow (it works through contingencies and requires sorting).

This slowdown was especially problematic for dask tables.

The new code does not currently support weights! Supporting them would require making a weighted variance computation. Should we go there? Done in #6205.

Requires #6204 and #6205.

Testing performance on 1000x1000 tables.

Before:

[normalize_only_parameters] with 5 loops, best of 3:
	min 233 msec per loop
	avg 235 msec per loop
[normalize_only_transform] with 5 loops, best of 3:
	min 120 msec per loop
	avg 121 msec per loop

After:

[normalize_only_parameters] with 5 loops, best of 3:
	min 71.7 msec per loop
	avg 73.3 msec per loop
[normalize_only_transform] with 5 loops, best of 3:
	min 117 msec per loop
	avg 119 msec per loop

The mean/variance estimation part is 3x faster. Of course, this PR has no influence on the actual transformation (the domain contains the same variables).

Includes

• Code changes
• Tests
• Documentation

[codecov]

Codecov Report

Merging #6202 (d212f84) into master (7d18600) will decrease coverage by 0.00%.
The diff coverage is 96.66%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #6202      +/-   ##
==========================================
- Coverage   86.66%   86.66%   -0.01%     
==========================================
  Files         316      316              
  Lines       68021    68027       +6     
==========================================
+ Hits        58951    58955       +4     
- Misses       9070     9072       +2     

[pavlin-policar]

Looks good to me.