Add regression tests for nullable_nonnegative decode filter in write_tables

[Copilot]

No test coverage existed for decode_columns mappings where the target column contains nulls and/or -1 sentinel values — the scenario addressed by the nullable_nonnegative filter added in this PR.

Changes

• activitysim/core/steps/_decode.py (new): Extracts _decode_output_column and _apply_decode_filter out of output.py into a pandas/numpy-only module, breaking the circular import chain (output → estimation → simulate → estimation) that prevented unit testing.
• activitysim/core/steps/output.py: Imports the two helpers from _decode.py; behavior is unchanged.
• activitysim/core/test/test_output.py (new): 7 focused unit tests covering all decode filter paths and the key regression scenario — a column mixing pd.NA nulls and -1 sentinels with nullable_nonnegative:

column = pd.array([0, pd.NA, -1, 3, pd.NA], dtype="Int64")
map_func, preserve_nulls = _apply_decode_filter(zone_labels, "nullable_nonnegative")
result = _decode_output_column(column, map_func, preserve_nulls=preserve_nulls)
# result: [100, NA, -1, 400, NA]