fix: use y.shape[1] in get_n_cols for Keras 3 compatibility (v0.1.4)

[gmgeorg]

Summary

• get_n_cols (pypsps/utils.py) used the old TF1-era y.get_shape().as_list()[1] for non-np.ndarray inputs. Under Keras 3, the tensor flowing through the nested CausalLoss -> OutcomeLoss call chain during graph-mode loss tracing is a KerasTensor, which has no get_shape() method at all -- raising AttributeError and crashing training with the default (posterior-weighted) loss.
• Fixed by using y.shape[1] unconditionally, which works uniformly for np.ndarray, eager tf.Tensor, graph-traced tensors, and KerasTensor, so the isinstance branch is no longer needed.
• Bumped pyproject.toml to 0.1.4 and added a CHANGELOG.md entry.

Test plan

• Added test_get_n_cols_np_array, test_get_n_cols_eager_tensor, test_get_n_cols_graph_mode_tensor, test_get_n_cols_keras_symbolic_tensor covering each tensor type get_n_cols needs to handle.
• Added test_toy_model_fits_in_eager_and_graph_mode, which fits build_toy_model with run_eagerly=True and run_eagerly=False.
• Verified against the pre-fix code: test_get_n_cols_keras_symbolic_tensor fails with AttributeError: 'KerasTensor' object has no attribute 'get_shape', confirming this reproduces the reported crash.
• Full suite passes: pytest pypsps/tests/ -> 89 passed.

🤖 Generated with Claude Code