refactor: GP predict method according

[mamarder]

Hello @dirmeier ,

I propose the following changes:

• Refactored GP predict method according C. E. Rasmussen & C. K. I. Williams, Gaussian Processes for Machine Learning, 2006, Chapter 2.2 Function-Space View, Algorithm 2.1. While this claims to be faster and numerically more stable the posterior covariance matrix was still not pos. def. in the examples/gaussian_process.py example which led to a failing cholesky operation in

      return distrax.MultivariateNormalTri(
          jnp.squeeze(mu_star), jnp.linalg.cholesky(cov_star)

  Note: Nevertheless this refactoring might contribute to issue https://github.com/ramsey-devs/ramsey/issues/11

• enabling 64bit precision for JAX library and adding jitter to the posterior covariance matrix led to a pos. def. covariance matrix

• having a pos. def. covariance function allowed adding confidence intervalls to the plot (see below)

[dirmeier]

Thanks for making the changes. This is great. I'll merge as soon as the tests are done.