Mud

experimenting with how to compute the map point, particularly with small standard deviations (0.1) and many measurements (over 500), revealed stability problems (overflow errors). this release gives the option to pass `log=True` to `map_problem`

support for vector-valued MUD solutions. no checks on max number of output dimensions, allowed to exceed input dimension, may cause uncertain behavior. may change in future to restrict to max equal to input dimension.

observed distribution already has `.prod(axis=1)` so support should be handled by default. needs tests.

now uniforms are set as the default, not normals.

implements MVP functionality for MAP point solution. Only difference in syntax between it and the MUD one is that a) you pass data as part of the likelihood definition, b) the second argument it takes has to have each column being a measurement. In the MUD one, we pass pre-transformed data.

implements sample-based solver with MVP functionality. This allows BET to no longer be required for solving the inverse problem.

Adding a DensityProblem class in mud.base which implements a bare-bones version of BET's discretization implemented in 2.3.0 but not ported to 3.0.0+