spirv-fuzz: Potential edge-case missed for TransformationDuplicateRegionWithSelection

A comment says:

We know that |entry_block| has only one predecessor, since the region is single-entry, single-exit and its constructs and their merge blocks must be either wholly within or wholly outside of the region.

And then asserts that the entry block only has one predecessor. I am not sure if this is right because control flow can converge early, so the single-entry, single-exit region could be within a selection construct and the entry block could be a "convergence block" within that selection construct (but not the header nor the merge block).

Start by locating the implementation of TransformationDuplicateRegionWithSelection and the assertion that the entry block has one predecessor. Review how convergence blocks and single-entry, single-exit regions are represented, then determine whether the reported control-flow case is valid. Done means the assertion is correct for the case or a regression test and appropriate fix cover it.

Written by the indexing model from the issue text.