Default to SVD for MvNormal in Latent GP conditionals

MvNormal was recently switched to default to use
cholesky decomposition which is very fragile for the
higher-dimensional situations encountered in `Latent` GPs.

Closes #7754

Author: michaelosthege

pymc/gp/gp.py

@@ -281,6 +281,7 @@ def conditional(self, name, Xnew, given=None, jitter=JITTER_DEFAULT, **kwargs):
         """
         givens = self._get_given_vals(given)
         mu, cov = self._build_conditional(Xnew, *givens, jitter)
+        kwargs.setdefault("method", "svd")
         f = pm.MvNormal(name, mu=mu, cov=cov, **kwargs)
 
         return f