fix: zero turnover limit, CAE/SAE minibatch edge cases, attention with no neighbors or long lag

[quantum-like]

Owning issue

Closes #55

Outcome

5 correctness fixes, one commit each, each with a focused test:

• A turnover limit of 0 liquidated the previous book. Zero now holds the previous weights; negative or non-finite limits raise. Test: tests/test_financial_correctness.py::test_zero_turnover_limit_holds_the_previous_book, ::test_negative_turnover_limit_is_rejected. (2a32358)
• CAE reported converged=True when hidden units were set and every minibatch had one row, so no step ran and the random initialization was returned. It now raises. The existing skip of a trailing singleton minibatch is unchanged. Test: tests/test_cae.py::test_cae_rejects_fits_that_never_update_parameters. (b0d261a)
• SAE could yield a one-row final minibatch, which crashed BatchNorm mid-fit. That tail is folded into the previous batch, and a sample with fewer than two observations is rejected. Test: tests/test_sae.py::test_sae_trains_when_the_last_minibatch_would_be_a_single_row, ::test_sae_rejects_a_single_training_row. (95557c0)
• Graph attention produced NaN outputs and gradients for a node with no available neighbor. It now falls back to self-attention. Test: tests/test_financial_correctness.py::test_graph_attention_stays_finite_when_a_node_has_no_neighbors, ::test_graph_attention_ignores_a_masked_only_neighbor. (7312029)
• Cross-sectional attention wrapped the slice when the lag exceeded the window. It now uses an all-pad key sequence. Test: tests/test_financial_correctness.py::test_cross_attention_lag_longer_than_the_window_does_not_wrap. (077948a)

Compatibility

• No public compatibility impact
• Compatibility impact is documented in the owning issue and migration guidance

A turnover limit of 0 now means "hold" rather than "liquidate", and invalid limits raise. Other changes affect only edge cases that previously crashed or produced NaN.

Verification

• Tests reproduce the prior failure or verify the requested behavior (each listed test fails on main and passes with its fix)
• Ruff lint and format checks pass
• ty passes
• Test suite passes: pytest tests/: 422 passed before, 431 passed after. ruff and ty were clean in the review environment.
• Package build passes when applicable (not run)
• Strict MkDocs build passes when documentation is affected (no documentation changes)
• Ecosystem qualification passes (not run)

Documentation and release

No documentation changes. The behavior changes above may be worth a line in the next patch-release notes.

Amir