bug: Zero mean function is trainable #499
Description
Issue #330 seems to have resurfaced after the switch to nnx. Thanks @chaozg for pointing this out. There was actually a test in place to stop this behaviour from resurfacing, but it seems to have been commented out during the refactor and never got included again.
The zero mean function should be fixed such that the constant remains zero, even after optimisation. At a first glance, this is likely achievable by setting the constant here to be a Static variable, with an alteration to the Constant mean function definition to allow it to be initialised with Static variables. The test for this should then be uncommented and updated to reflect the changes made when switching to nnx.