Add user-supplied domain interior function and fix float roundoff bug in gp-1d demo notebook

[asher-m]

This PR:

1. Adds support for a user-supplied domain interior function in UncertainSCI.gp.GaussianProcess.get_sample_point, and
2. Fixes a float32 (default precision in JAX) roundoff bug that caused sample points to appear to not be optimized
   • Specifically, when the gradient of the variance w.r.t. the coordinate is too small, float error would round this to zero resulting the sample point remaining "stuck in place" during gradient ascent.
   • This is not a genuine fix in the normal sense, more of a work-around. Unfortunately, certain "symmetric" decisions in the choice of how to "densely" sample the coordinate space to find an initial sample point from which to start gradient ascent can get the optimizer stuck at a zero-gradient non-maximum.
   • In the notebook, I fix this by uniformly randomly sampling the coordinate space, rather than repeatedly using the same (uniformly) spaced sample points to start gradient ascent.