Fix CausalForestDML ATE and ATT standard errors for multiple outputs

[kbattocchi]

Summary

Fixes #1066.

_CausalForestFinalWrapper._ate_and_stderr previously counted all non-missing elements of the doubly robust prediction tensor, rather than counting observations separately for each outcome-treatment pair. This understated ATE and ATT standard errors whenever there were multiple outcomes or treatment contrasts.

• Count non-missing predictions along the sample axis (axis=0), including after ATT masking.
• Preserve point estimates and output shapes, and document the per-pair observation counts.
• Add deterministic regressions for binary and three-arm treatments; vector, single-column, and multi-output outcomes; missing rows and unequal column missingness; and masked subsets.
• Check fitted-estimator ATE/ATT standard errors, confidence intervals, z-statistics, and p-values, including every treatment group.

The separately identified multi-output model-selection scoring issue is intentionally out of scope.

Validation

• Before the correction: 24 deterministic regression cases failed and 12 single-output/single-contrast controls passed.
• After the correction: all 42 new regression cases and the existing causal-forest ATE inference test passed (43 total).
• Ruff passed for both changed Python files; git diff --check passed.

Local fitted-estimator validation initially encountered a native-code crash with Cython 3.3.0. Rebuilding with the repository's last-known-good Cython 3.2.4 resolved it; the passing results above use that build.