Avoid repeated (kernel) computation when optimizing classical hyperparameters

[rothma]

[David-Kreplin]

In case you are using projected Quantum Kernel, the results are cached automatically.

[rothma]

but not for fidelity kernels