Optimiser more lightning-centric

[felix-e-h-p]

Relating to #487

Proposed changes outlined as follows:

• All schedulers now returned via Lightning style dict with monitor, so Lightning steps them automatically on logged val metric.
• LR is owned by the wrapper (public lr), so LR finder / tuning connects more cleanly and no need for model.lr.
• The optimiser wrappers return {"optimiser", "lr_scheduler": {"scheduler", "monitor"}}.
• Interface is standardised to AbstractOptimizer.call(model), so every wrapper takes a model and returns an optimiser config.
• Less coupling (no model.lr), no mixed return types, and fewer custom hooks whereby Lightning drives ReduceLROnPlateau.

Not strictly necessary, this is a Lightning-centric proposal really and definitely open to other ideas / options!