Bug fix: stop using dangling ptrs and buildPattern() in fixPartitions()

[StefanFlaumberg]

This pull request fixes two issues in the fixPartitions() function from phylotesting.cpp:

1. ModelOMatic can change some subalignments effectively through the stree->at(i)->aln pointers. This means that the respective saln->partitions[i] pointers remain dangling, as they cannot be reset in place due to the general code design. Comparison with a dangling pointer has an implementation-specific behaviour in c++, it might work or might not.
   Therefore, I opt to avoid it here simply by unconditionally resetting all saln->partitions pointers and relevant saln info (see below). This means that now the reset happens for each PartitionFinder call, but since it happens only once, the time overhead is negligible.
2. After my recent alignment refactoring, SuperAlignment::buildPattern() expects that the alignment pattern vector is empty. This guarantees that the function is used only during superalignment initialization, as any other usage would likely be incorrect.
   Here, in fixPartitions(), we change sequence types of subalignments (and hence their pattern numbers and contents), but the patterns of taxon presence/absence remain the same. So saln->taxa_index remains the same as well and it would be pointless to call saln->buildPattern(), even if we could. But since the pattern contents of the subalignments might have changed, we have to call saln->computeConstSites() and saln->orderPatternByNumChars here.

[bqminh]

looks good