[InstCombine] use m_APInt to allow ashr folds for vectors with splat constants

We may be able to assert that no shl-shl or lshr-lshr pairs ever get here
because we should have already handled those in foldShiftedShift().

llvm-svn: 292726