`LC2ST_NF` does not honor the `LC2ST` method contract

LC2ST_NF subclasses LC2ST, but its evaluation methods have a different contract. In LC2ST, theta_o are posterior samples at x_o. In LC2ST_NF, evaluation runs in the flow's base space on samples drawn once in __init__. The class docstring says so: no theta_o is passed to the evaluation functions.

The five overrides (get_scores, get_statistic_on_observed_data, p_value, reject_test, get_statistics_under_null_hypothesis) therefore drop theta_o. Pyright flags them under reportIncompatibleMethodOverride, see #1979.

A signature patch does not fix this. #1984 tried it: theta_o becomes an ignored parameter, x_o needs a None default plus a guard, and positional x_o calls break. We dropped that part of the PR.

The fix is structural. I see two options:

• Composition: LC2ST_NF holds an LC2ST and exposes its own (x_o, ...) API.
• A shared private base for the training and null machinery. LC2ST and LC2ST_NF each define their own public evaluation methods on top.

Both change the public hierarchy, isinstance(lc2st_nf, LC2ST) stops being true, so this needs a deprecation note.

Start by reading the LC2ST and LC2ST_NF evaluation methods: get_scores, get_statistic_on_observed_data, p_value, reject_test, and get_statistics_under_null_hypothesis. Compare their contracts and the Pyright error from #1979, then choose a hierarchy or composition design that preserves each public API. Done means the overrides are compatible, positional x_o calls remain valid, and the isinstance change has a deprecation note.