MAINT: Merge some 1D and 2D methods into ND mixin methods

[derb12]

Description

Internally refactored the _Algorithm and _Algorithm2D _register (now _handle_io) wrappers and _setup_* methods to have the same signatures so that the simpler 1D and 2D methods could be generalized into ND mixin classes. This allows having just one canonical method for an algorithm rather than 2. The ND mixins are simply inherited to provide the methods to the appropriate _Algorithm[2D] subclasses.

Also refactored the PenalizedSystem, PSpline, WhittakerSystem2D and PSpline2D classes to have the same method calls so that Whittaker smoothing and P-spline smoothing could be generalized under a common penalized least squares (pls) setup. Combined with the above, this allows one method to handle 1D and 2D Whittaker smoothing and P-spline implementations, ie. one method instead of four methods.

With the above changes, this PR condenses 48 separate methods into just 17. This will be a huge help for future maintenance or changes, such as adding masking in #72, since only one code path needs updated rather than 2 or 4 for a method. Note that most of the smoothing and morphological algorithms are also low hanging fruit for this ND generalization, but the use of padded convolution in 1D makes it a bit harder to generalize (and seems quite excessive upon review). In the future, I need to see how bad edge effects are if I replace the padded convolution with just a convolution; or add a 2D padded convolution.

Tests should pick up if I screwed up majorly somewhere (eg. 2D input gets called to 1D method or a Whittaker method uses p-splines). I also generated local integration test data using this gist based on the development branch and ran the comparison tests in this gist with this PR's changes as a sanity check to ensure the 1D and 2D methods retained their output, at least against changing their main input parameters.

Type of Pull Request

• Bug Fix
• New Feature
• Miscellaneous Changes (refactor, code improvements, etc.)
• Documentation or Example Programs

Pull Request Checklist

• New code and/or documentation is valid for use with the BSD 3-clause license.
• New code is fully documented with docstrings that follow Numpy style,
  if applicable.
• New code follows PEP 8 standards as closely as possible, if applicable.
• Added/updated tests and ensured they pass locally, if applicable.
• Verified that documentation builds locally, if applicable.
  • viewcode links no longer get generated for the 2D methods that directly inherit from the ND mixins. Will see if it's fixable in a follow up, but it's not a big issue if not.

[derb12]

Not related to the changes in this PR, but I'm noticing the test tests/test_whittaker.py::TestAsPLS::test_avoid_overflow_warning is emitting (intended) warnings in the CI jobs (again not seeing it locally though...). I should filter out that warning like in the similar TestAirPLS test case. And probably run tests with -W error like I do on the nightly tests for all but the min dependency case so I catch stuff like this more easily; it looks like this warning's been getted emitted as far back as the CI logs go.

[derb12]

Looking at the docs build, my local integration tests should've been for all methods rather than just the ones converted to ND... In commit cd0bd07, jbcd should have converted to direct_solve instead of solve since the rhs values are not multiplied by the diagonals added to the lhs, so those diagonals shouldn't have been considered as weights. Reverting to using direct_solve fixes jbcd's output, and the updated local integration tests that look at ALL methods now pass.

CI won't detect the fix, so I'll just merge this and make the change in the dev branch.

Also unrelated to this PR, but adaptive_minmax's fit in the docs is also wrong compared to the stable docs, but that goes back to 5d0f3a7. The correct fix for #62 should have been to just mask out weights == 0 or weights < eps rather than use the full weighted residual. Or maybe change adaptive_minmax to use Lagrange multipliers like its literature implementation instead of heavy weighting. Will follow up after merging. Need to add at least basic integration tests so these sort of bugs don't occur...