remove redundant PieceWiseLinearGP add ExactW Kernels

[Jimbo994]

Motivation

With the advent of the engineered features, we can now, in principle, get rid of the PieceWiseLinearGP because it can be fully recreated using the InterpolateFeature engineered feature. So, for future maintenance, it is better to remove the now-redundant PieceWiseLinearGP and accompanying code from, for instance, torch_tools.py and other locations.

In the meantime, I also worked on exact implementations of the WassersteinKernel. The WassersteinKernel was computed on an interpolated grid. The ExactWassersteinKernel considers the union of unique x-values and then uses the trapezoidal rule, together with identifying the sign of break points, to perform the computation exactly. Both W1 and W2 distances are supported. The W1 implementation matches the approximate WassersteinKernel at an adequately high number of interpolation points.

The union of unique x-values can become quite large when having high-dimensional problems, a large number of raw samples/restarts in the acq. func. optimization, etc. Which on moderate compute (e.g., small k8 compute), can lead to memory issues. Therefore, there is also some chunking in the code so that we loop over the union. For CPU this was also slighly faster. Thoughts on this implementation @jduerholt?

Lastly, fixed a bug in the interp1d, which could lead to gradfNone errors due to division by zero. In principle, this should never happen, as inequality constraints are set, but initial candidates provided by Botorch sometimes violate these.

Have you read the Contributing Guidelines on pull requests?

Yes

Have you updated CHANGELOG.md?

Yes

Test Plan

Removed old tests.
Created new tests for new features.
Ran tests.

[jduerholt]

Looks really, really good to me. I love it. Only some minor comments.

[Jimbo994]

How about this @jduerholt

[jduerholt]

Looks good to me. Thanks!