fix(predicate): resolve _ROW_ID by name, not by its placeholder leaf index

[plusplusjiajia]

Purpose

_ROW_ID is not a table column, so PredicateBuilder cannot resolve it and callers hand-build the leaf with an index of their own — 0, as the existing test_with_filter_extracts_row_id_ranges does. That index is a placeholder, but Rust binds predicate leaves by index at construction and never rebinds by name, so every
index-based resolution binds the predicate to whatever field sits at that index. On a table whose first field is a partition key, WHERE _ROW_ID >= 10 pruned partitions by comparing them against row ids; on a primary-key table it was pushed down as a key conjunct and dropped key versions before the merge; a residual
_ROW_ID != 5 filtered the first user column.

Java rebinds by name (PredicateRemapper, and rowIdSafeResidualFilter for row ids); pypaimon's _change_index looks the field up by name. Rust was the only implementation with no name-level step.

Brief change log

Resolve by name. spec::is_row_id_column is the single definition, and every positional resolution consults it first: the three Predicate index-mapping primitives (which is what keeps a row id off partition keys, bucket keys and PK pushdown), residual mask evaluation, per-file remapping, scan widening, the projected
field-id set, Parquet/ORC decoder pushdown, stats pruning, and the data-evolution BLOB dependency check. DataEvolutionReader attaches _ROW_ID when a residual needs it but the caller did not project it.

Row ranges. extract_row_id_ranges returns Some(vec![]) for an empty union rather than None — "matches no rows" and "nothing extracted" are different, and reporting the latter let an enclosing AND skip the subtree and overstate its ranges. Ranges are only a superset, so ranges_represent_conjunct lets a conjunct leave the
residual only when the ranges stand in for it exactly, and row_ranges_from_filter clears a derived range when its filter is replaced.

Fail closed. Only data-evolution reads can synthesize _ROW_ID; elsewhere the predicate is unenforceable and every alternative is a silent wrong answer, so reject_row_id_filter rejects it on the read path (not the builder — TableRead is public): primary-key merges, format tables, Vortex's scanless fast path, and the Mosaic
and Blob readers. This is only for _ROW_ID — it is the only reserved name a predicate references, and the only one with no schema position at all.