Global_step workaround

[saschafrey]

The trick of creating a global_step property that is manual incremented does not increase the global_step of the trainer which is used by the ModelCheckPoint callback (Lightning-AI/pytorch-lightning#17167). A workaround is simply to call a step on a dummy optimiser with self.optimizers().step() inside the training_step method.

self.optimizers() returns a wrapped version of the optimiser(s) in configure_optimizers(), as this is 'None' in the examples given (due to manually doing opt with whatever jax optimiser) it does nothing but increase the step counter.