[ENH] Feature Constructor: Vectorize when possible

[janezd]

Issue

Resolves #5911.

Description of changes

• Try to compute the function on columns; fall back to per-instance if this fails.
• Import functions from np instead of from math when available.
• Show all functions in combo (avoid compiling a separate list of functions)
• Remove cumsum, cumprod, which created two-dimensional data instead of a column

Includes

• Code changes
• Tests

[codecov]

Codecov Report

Merging #5949 (7a2b9e8) into master (647f5da) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master    #5949   +/-   ##
=======================================
  Coverage   86.28%   86.28%           
=======================================
  Files         315      315           
  Lines       66906    66920   +14     
=======================================
+ Hits        57728    57745   +17     
+ Misses       9178     9175    -3     

[ajdapretnar]

This works for my limited amount of use cases.

[markotoplak]

I was looking at math vs numpy docs and did not see any cases where I would predict conflicts. That does not mean they aren't any.

@janezd also raised the question about import everything from math. I would rather name a subset of functions, perhaps those that are already available in the python 3.10, so that we do not break any current workflows. Imports should be "soft" since not all of them are available on oldest supported python.

[janezd]

@markotoplak, I checked the functions in math and all seem useful, so I'd keep it the selection it is.

As for reporting warnings and errors in expressions: the problem here is that the user can write an expression that cannot be evaluated by numpy, in which case we fall back to per-instance evaluation. To be on the safe side, I'd fall back at any sign of trouble, that is, at any warning or error. When this happen, the execution will be slower (that is: as slow as before this PR) and any errors will be properly reported.

[ales-erjavec]

The random functions (gauss, normalvariate, ...) no longer work on per instance basis. Single generated value is used for the whole column (e.g. sepal_length + gauss(0, 1) or just gauss(0, 1))

[janezd]

Closed in favor of #5975.