getNumberOfTrees: report the support, not the graph

[alexeid]

ITreeDistribution.getNumberOfTrees() was documented only as "the number of trees (topologies) in this distribution", which is ambiguous between the CCD graph and the distribution's support. They coincide for CCD0/CCD1 but not for full-support models.

KRegCCD and MRegCCD both override containsTree to return true, yet inherited the graph-based count from AbstractCCD. On a 129-taxon posterior KRegCCD reported log 88 against a true support of log 582, contradicting its own containsTree.

• document the contract on the interface: support, not graph
• override in KRegCCD and MRegCCD
• move numberOfRootedTopologies and logBigInteger to AbstractCCD so all models share them; UniformEscapeCCD.numberOfRootedTopologies still delegates
• delete a commented-out logBigInteger in CredibleCCDComputer that was incorrect (new BigDecimal(bigInteger).scale() is always 0, so it reduced to Math.log(val.doubleValue()))
• add NumberOfTreesContractTest

Note the double-overflow boundary is ~155 taxa: 129 taxa needs 840 bits and converts fine, 160 needs 1093 and overflows. Coal160/Coal320/Yule200/Yule400 are past it.

Existing callers (NNICladeExpansion, RogueDetection) use CCD0/CCD1 where graph and support coincide, so are unaffected. Full suite: 110 tests, 0 failures.