Make operator target acceptance probability configurable; bump 1-D defaults to 0.3

[alexeid]

Summary

Adds a targetAcceptanceProbability XML input on the Operator base class so users can override the auto-tune target per operator without code changes.
Refactors the resolution path (input -> getDefaultTargetAcceptanceProbability() -> base default) and migrates every existing override to the new method name.
Adjusts a few defaults at the same time.

What changes

Base class (Operator.java):

• New Input<Double> targetAcceptanceProbabilityInput (default Double.NaN, meaning "use operator default").
• getTargetAcceptanceProbability() now returns the input if set, otherwise calls getDefaultTargetAcceptanceProbability().
• New getDefaultTargetAcceptanceProbability() returning 0.234, which subclasses override.

Migrated overrides (functionally equivalent, just renamed): BactrianScaleOperator, BactrianNodeOperator, BactrianRandomWalkOperator, BactrianIntervalOperator, BactrianDeltaExchangeOperator, EpochFlexOperator, AdaptableVarianceMultivariateNormalOperator (legacy and spec, drops final), IntervalScaleOperator (spec), AbstractScale (spec), RealRandomWalkOperator (spec), DeltaExchangeOperator (spec), IntervalOperator (spec).

Default bumped 0.234 -> 0.3 (1-D operators that were inheriting the base default):

• SubtreeSlide, BactrianSubtreeSlide
• ScaleOperator
• TipDatesScaler, TipDatesRandomWalker (BactrianTipDatesRandomWalker inherits via its parent)
• RealRandomWalkOperator (legacy inference)
• DeltaExchangeOperator (legacy inference)

Default reverted 0.3 -> 0.234 (multi-D operators that should use the asymptotic target): BactrianUpDownOperator, UpDownOperator (spec).
The legacy inference/operator/UpDownOperator.java already inherited 0.234 and is unchanged.

Operators with deliberate non-default values are kept (EpochFlexOperator 0.4) but renamed to getDefaultTargetAcceptanceProbability.

Why

The 0.234 default comes from Roberts, Gelman, Gilks (1997) and is the asymptotic optimum for random-walk Metropolis as dimension increases.
For 1-D moves the optimum is closer to 0.44 (Gelman, Roberts, Gilks 1996; Roberts and Rosenthal 2001), so 0.234 is too low.

We observed this in practice on a real-data analysis: starting BactrianSubtreeSlide at size=0.002 with optimise="false" gave ~53% acceptance, but with optimise="true" the auto-tuner repeatedly increased step size to drive acceptance toward 0.234, and after enough moves the 1/n Robbins-Monro gain became too small to walk the size back when steady-state acceptance fell to ~5%.
A target of 0.3 is a conservative compromise that does not require the tuner to push as hard against the structural acceptance ceiling of these 1-D operators.

The XML input lets users override on a case-by-case basis without further code changes:

<operator id="treeSlide" spec="...BactrianSubtreeSlide" tree="@psi" weight="9.25"
          targetAcceptanceProbability="0.4"/>

Test plan

• mvn -pl beast-base -am compile clean
• mvn compile (all modules) clean
• mvn -pl beast-base test -Dtest='*Operator*Test' -> 20/20 pass
• Reviewer to confirm no XML schema/template assumes the absence of this input
• Re-run the affected XML with the new defaults and confirm BactrianSubtreeSlide auto-tunes to a sensible step size