Extend native Calc function coverage with released Flink semantics

[liuyongvs]

Runtime SQL using common exact numeric, string, binary and collection functions previously fell back at Calc. This PR adds the verified overloads using released Flink expression generation through the existing batch JVM bridge, preserving result types, values, NULL behavior and error evaluation. BOOLEAN IF uses the existing lazy native CASE path.

The branch contains one commit for each issue:

Issue	Added coverage
#208	Exact unary minus, ABS/SIGN, integral FLOOR/CEIL/TRUNCATE
#209	STRING, mixed exact numeric and TIMESTAMP/LTZ extrema
#210	BOOLEAN IF results
#211	BOOLEAN to STRING/CHAR/VARCHAR casts
#212	LIKE ESCAPE and SIMILAR TO
#213	REGEXP and scalar regex helpers
#214	PARSE_URL
#215	PRINTF
#216	Dynamic BTRIM/LTRIM/RTRIM sets
#217	FROM_BASE64, numeric SPLIT_INDEX, dynamic SHA2 and classification
#218	Binary STARTSWITH/ENDSWITH/ELT
#219	REGEXP_EXTRACT_ALL and STR_TO_MAP

Generated expressions keep intermediate computations and sensitive UTF-16 consumers together. Collection results use the whole-Calc JVM callback with owned Arrow outputs and filter-before-projection evaluation. Base64-decoded STRING results can contain arbitrary bytes: final scalar outputs use Arrow Binary and the row transpose preserves those bytes; they cannot feed a native columnar sink as Utf8. No Rust/JNI ABI or third-party dependency changes are required.

Admission remains conservative. Runtime string extrema and dynamic trim sets require proven Java-backed external DataStream inputs; unknown or binary-backed representations stay on Flink. Sensitive character results crossing relational operators and unverified complex boundaries retain fallback. DECIMAL FLOOR/CEIL and FLOAT/DOUBLE TRUNCATE are outside this scope. The operator remains columnar, while generated functions execute on the JVM.

Validation on canonical main 89da01cd, JDK 17 and released dependencies:

• Flink 2.2.1: 268 targeted integration/planner/bridge tests pass, including the twelve issue suites, JSON identity, temporal and UDF regressions, native Parquet boundaries and allocator checks.
• Re-ran the same 562 runtime-source SQL cases used for the issue baseline: 525 native passes, 6 fallback passes, 31 pre-existing both-error cases, zero result/schema mismatches and zero native-only failures. All 305 newly admitted successful cases move from fallback to native. The six remaining fallbacks are floating-point TRUNCATE controls. Both-error cases are not counted as exception-equivalence proof.
• Focused tests separately assert host/native root exception class, message and evaluation phase; exercise multi-batch NULL/Unicode/binary cases, invalid regex/format/escape/hash/Base64 inputs, short-circuiting and all-filtered batches. Raw decoded bytes are checked at the RowData sink.
• Flink 1.18.1: 24 selected tests pass; 11 cases are explicitly N/A because the release lacks those SQL functions. Additional nullable-map boundary checks pass on both lines.
• Release-only diagnostics use 200,000 rows, 64-byte payloads, parallelism 1, one warmup and three measured runs, matched-source controls and both transposes. The 13 representative expressions reach 0.630x–0.913x stock Flink throughput in this short row-fed workload. This is coverage work and does not claim standalone scalar or full-pipeline speedups; detailed results and retained gates are in docs/operators/calc-filter.md.

Closes #208
Closes #209
Closes #210
Closes #211
Closes #212
Closes #213
Closes #214
Closes #215
Closes #216
Closes #217
Closes #218
Closes #219