Make disconnect() symmetric with connect(), and cover two remaining wrapper sites

[alexeid]

Follow-up to #132 (review findings), targeting mcmc so it lands inside that PR rather than separately. See #131.

1. disconnect() never got the inner-"mcmc" redirect

connect() gained the redirect in f6a369c; disconnect() did not. scrubAll() calls the two as the branches of a single decision:

} else if (connector.isActivated(...)) {
    connect(connector, context);
} else {
    disconnect(connector, context);
}

So with a wrapper Runnable, objects get connected to the inner MCMC but are never removed from it again: disconnect() resolves the wrapper, target.getInput(inputName) throws IllegalArgumentException, and the surrounding catch-and-print swallows it. The failure is silent — toggling a checkbox off, deleting a partition, or switching a clock model leaves stale operators/loggers on the inner MCMC and the generated XML is wrong rather than crashy.

Extracted the redirect as redirectToInnerMCMC() and used it from both, with a javadoc note that it must stay used by both.

Also gave disconnect(BeautiConnector, PartitionContext) the targetID='mcmc' -> mcmc.get().getID() translation connect(BeautiConnector, ...) already has. This is a no-op for the shipped templates — Standard.xml declares <run spec="beast.base.inference.MCMC" id="mcmc">, so the translation resolves to the same string — but it matters for a wrapper template whose runnable carries a different ID.

2. Two remaining sites that assume MCMC's inputs

These read MCMC inputs off doc.mcmc.get() without a cast, so the (MCMC) sweep didn't catch them, but they fail the same way — getInput()/setInputValue() throw IllegalArgumentException for an unknown input name (BEASTInterface.java:413):

• OperatorListInputEditor:66 — getInput("operatorschedule"). The Operators panel is in BEAUti throws ClassCastException in multiple panels when the active Runnable wraps MCMC (e.g. PathSampler) #131's repro list, so this one was still broken.
• MRCAPriorInputEditor:329/354 — setInputValue("operator", ...) and getInput("operator"), i.e. the tip-date sampling toggle.

Both routed through doc.getMCMC(), following the null-guard style already used elsewhere in #132.

Verification

• mvn -o -pl beast-fx -am -DskipTests compile clean.
• Headless BEAUti tests: BeautiMRCAPriorTest 1/1 pass, LinkUnlinkTest 14 run / 0 failures / 7 skipped (pre-existing @Disabled). BUILD SUCCESS.

Not addressed here

Null handling across #132 is still split roughly half-and-half — BeautiDoc:1595/:1641 (getMCMC().posteriorInput.get() in scrubAll/setUpActivePlugins) and the deepCopyPlugin call sites will NPE where other sites carefully guard. That's a contract decision for you: either those are bugs, or the guards are dead code and getMCMC() should throw a named IllegalStateException instead of returning null. Left alone deliberately so this PR stays reviewable.