Allocate optimizer state afresh on every rule invocation

[jessegrabowski]

A rule such as adam(1e-3) now allocates fresh optimizer state on every invocation, and the updates dict it returns is the only state-bearing object. reuses_state, the decorator that memoized a rule's buffers across invocations through a ContextVar, is deleted along with the registry behind it. Two training functions that share momentum are compiled from one dict, and compile_train accepts that dict in place of a rule.

Every learning_rate argument now accepts a schedule, in the *_updates functions and in scale as well as the aliases. A rule given a schedule reads it off the clock it already counts its own steps on and writes that clock into its updates, so a scheduled rule holds one clock and a checkpoint taken from the dict resumes the schedule where it left off. scale_by_schedule stays as the optax-named alias of scale.

skip_if now holds a rule's clock back through a skipped step along with the rest of the rule's state. A clock that advanced anyway bias-corrected the next step against a moment that never saw it, which damped adam's first step after a skip by about half. A clock the caller holds and passes into a schedule is not the rule's state and keeps counting.

counter is removed in favor of step_counter. The other commits close gaps this contract exposed: a rule drops the gradients it does not descend, compile_train refuses a dict that leaves a collected parameter unwritten, nonfinite checks optimizer state and not only parameters, reduce_on_plateau refuses a scale no step reads, and guard and policy counters are int64.

Closes #160

---

📚 Documentation preview 📚: https://pytensor-ml--163.org.readthedocs.build/en/163/