Add predict with mean only option

[sgreenbury]

As mentioned in #638 (review), in some cases only a prediction of the mean is required even when working with ProbabilisticEmulator.

Adding a new method to Emulator base class that returned the mean (TensorLike) could be implemented. Skipping covariance calculations would probably need to be added into the API of predict().

[radka-j]

This will probably become useful in MCMC calibration. Running the same example with an MLP model is a LOT faster than with the GP right now. I used the Epidemic simulator and run it for 2 chains for 1250 samples. With the MLP those took 15 and 18 seconds. With the GP the two chains took 1 minute 2:40 minutes.

[sgreenbury]

From discussion with @radka-j, this also relates to the potential effects of make_positive_definite or other transforms. For transforming the mean only we can provide guarantees here where covariance restructuring we might not be able to.