Reintroduce argument caching for float stepsizes

[sbanert]

In pull request #1166, argument caching for proximal operators was deactivated, since lists are not hashable. In order to avoid performance regressions, this should be reintroduced for scalar stepsizes.

The commit where the caching was removed is the following: https://github.com/odlgroup/odl/pull/1166/commits/cfcbaa0cdbc2ef14df0700839c5b524f1313da04