fix(table): honor physical schemas in raw evolution reads

[JingsongLi]

Summary

• resolve the physical fields of each raw-convertible data-evolution file from its schema ID and write_cols
• push predicates into raw-convertible reads when every row-id segment has a single file provider
• preserve partial-column and missing-predicate-column NULL semantics by remapping predicates against the physical file schema
• keep _ROW_ID, transformed BLOB/BLOB-view predicates, and column-merge reads on the residual-only path

Why

Data-evolution reads previously suppressed all file-level predicates because overlapping files may provide different columns for the same logical rows. That restriction is unnecessarily broad for raw-convertible splits: when each row-id segment has only one file, no column-wise merge is required, so the file reader can safely apply predicate pruning and row filtering. The exact post-read residual remains as a format-independent correctness backstop.

Raw-convertible files must also be decoded using their physical write_cols, even when their schema ID matches the table schema. This is required for positional formats such as .row. For example, if the table schema is [BLOB payload, INT id] but a data file physically stores only [id], decoding it as the full table schema interprets the id bytes as a BLOB offset and can read past the row payload. Resolving the physical fields fixes that pre-existing issue and also gives predicate pushdown the correct schema mapping.

Validation

• added test_raw_row_file_uses_write_cols_when_blob_precedes_projection; it reproduces the positional decode failure with the old field resolution and passes with this change
• cargo fmt --all -- --check
• cargo test -p paimon data_evolution_reader::tests::test_raw_row_file_uses_write_cols_when_blob_precedes_projection -- --nocapture
• cargo test -p paimon data_evolution_reader::tests::test_evolution_ -- --nocapture
• cargo clippy -p paimon --lib -- -D warnings
• git diff --check