Nested-column conjuncts are silently dropped from the row filter #21

Same class as #19, different drop path, still open after #20.

pushdown_columns (src/datafusion/src/reader/plantime/row_filter.rs) returns None for three reasons: an empty column set (fixed in #20), non_primitive_columns, and projected_columns. build_row_filter then .flatten()s those candidates away. Because DataFusion has already removed the FilterExec, the liquid scan is the only place the predicate is applied, so a dropped conjunct widens the filter.

Repro — 8 rows, st is struct<a int>, one row has st.a = 3:

SELECT id FROM t WHERE id >= 0 AND st.a = 3

DataSourceExec: ... projection=[id], file_type=liquid_parquet   <-- no FilterExec
-> 8 rows, expected 1

Fails on both cold and warm cache. Upstream DataFusion 54 supports struct-field-access pushdown (struct_field_accesses in its own PushdownChecker), so it plans the predicate as fully pushed while the liquid fork of row_filter.rs drops it.

The honest fix is to refuse the pushdown when a conjunct cannot be evaluated, so a FilterExec is retained, rather than dropping it. The doc paragraph on build_row_filter that sanctions dropping conjuncts ("the returned RowFilter may not contain all conjuncts ... will be ignored") encodes the assumption that #19 disproved and should go with it.

Found by adversarial review of #20.