Fix crash when initializing a WordAlignmentMatrix with one dimension

[pmachapman]

This PR fixes a crash exposed by the e2e test for sillsdev/machine.py#336, where the target pre-translations were blank.

Interestingly, the crash only occurs on Linux and macOS: https://github.com/sillsdev/thot/actions/runs/31657210462

The training data that exposes the crash is at crash_data.zip, and can be run with a unit test along the lines of:

TEST(SymmetrizedAlignmentModelTest, crash)
{
  chdir("/home/peter/machine/docker-compose/builds/6a7be055db58e46ea2adfd2f/model");
  auto direct = make_shared<FastAlignModel>();
  direct->load("src_trg_invswm");
  auto inverse = make_shared<FastAlignModel>();
  inverse->load("src_trg_swm");

  auto* model = new SymmetrizedAlignmentModel(direct, inverse);
  model->setHeuristic(SymmetrizationHeuristic::GrowDiagFinalAnd);

  for (int i = 0; i < 74; ++i)
  {
    WordAlignmentMatrix waMatrix;
    LgProb prob = model->getTrainingAlignment(i, waMatrix);
    EXPECT_EQ((double)prob, (double)SMALL_LG_NUM);
  }
}