Support Gaussian latent state

[fritzo]

The idea is to allow slowly-varying latent Gaussian state of some dimension (e.g. 3).

• Factor out TreeTrainer from TreeCatTrainer, to make structure learning reusable.
• Implement a TreeGaussTrainer (in progress).
  • Choose sufficient statistics.
  • Implement dynamic programming algorithm in treegauss_add_row()
  • Work out math for logprob() and compute_edge_logits().
• Implement data generation for tests.
• Implement TreeGaussServer.
• Implement real data import in format module.