test: comprehensive causal_prediction tests with synthetic data and PR #1371 regression coverage

[Copilot]

The existing causal_prediction test suite had 14 mostly structural tests with no coverage of the core regularization logic, CACM training, or the bugs fixed in PR #1371. This PR expands coverage to 51 tests using synthetic tensors.

Shared fixtures

• gaussian_reg — pre-built Regularizer(E_conditioned=True, gaussian, gamma=1.0); auto-skips when torch absent
• tiny_dataset_factory — factory for torch.utils.data.Dataset integer sequences

New coverage

• _split_by_attribute — two-label split, single-label passthrough, three-label
• _optimized_mmd_penalty — correctness vs. naive for gaussian (2- and 3-tensor) and linear kernels; empty-tensor skip; float64 small-gamma precision
• unconditional_reg — all four branches (E_eq_A × E_conditioned), both optimization modes
• _compute_conditional_penalty — positive/zero penalty cases, value-based grouping, optimization equivalence
• conditional_reg — all branches, optimization equivalence, cross-env merging
• CACM — attr_types=None / E_eq_A=None defaults, no mutable-default aliasing, training_step with all attr types
• AdamW — base-class acceptance, configure_optimizers return type
• ERM — end-to-end training_step with synthetic data
• Dataloader — split_dataset size, reproducibility, seed sensitivity

PR #1371 regression tests

Test	Bug covered
test_compute_conditional_penalty_value_based_grouping	Tensor-as-dict-key: equal-valued tensors from different envs treated as distinct keys → inflated penalty
test_unconditional_reg_not_e_conditioned_same_attr_gives_no_penalty	Old code computed spurious cross-env MMD when both envs share one attribute value
test_conditional_reg_not_e_conditioned_cross_env_merging	Cross-env sample merging in conditional_reg
test_cacm_e_eq_a_none_does_not_raise	attr_type_idx in None raised TypeError with old E_eq_A=[] default
test_optimized_mmd_penalty_fp64_precision_small_gamma	Float32 vanishing penalty for gamma ≤ 1e-6; fixed by float64 accumulation