[SPARK-40610][SQL] Support unwrap date type to string type

[wangyum]

What changes were proposed in this pull request?

This PR enhances UnwrapCastInBinaryComparison to support unwrapping date type to string type.

Why are the changes needed?

Avoid always fetching all partitions because the partition filters cannot be pushed down to the Hive metastore. For example:

CREATE TABLE t1(id int, dt string) using parquet PARTITIONED BY (dt);
EXPLAIN SELECT * FROM t1 WHERE dt > date_add(current_date(), -7);

Before Spark 3.0. It pushes partition filters to Hive metastore:

== Physical Plan ==
*(1) FileScan parquet default.t1[id#2,dt#3] Batched: true, Format: Parquet, Location: PrunedInMemoryFileIndex[], PartitionCount: 0, PartitionFilters: [isnotnull(dt#3), (dt#3 > 2023-02-27)], PushedFilters: [], ReadSchema: struct<id:int>

After SPARK-27638. Because it can not convert partition filters to hive metastore filters, it will not push partition filters to Hive metastore. As a result, it always takes all the parititons:

== Physical Plan ==
*(1) ColumnarToRow
+- FileScan parquet default.t1[id#5,dt#6] Batched: true, DataFilters: [], Format: Parquet, Location: InMemoryFileIndex(0 paths)[], PartitionFilters: [isnotnull(dt#6), (cast(dt#6 as date) > 2023-02-27)], PushedFilters: [], ReadSchema: struct<id:int>

After this PR. It unwraps date type to string type and then pushes partition filters to Hive metastore:

== Physical Plan ==
*(1) ColumnarToRow
+- FileScan parquet spark_catalog.default.t1[id#0,dt#1] Batched: true, DataFilters: [], Format: Parquet, Location: InMemoryFileIndex(0 paths)[], PartitionFilters: [isnotnull(dt#1), (dt#1 > 2023-02-27)], PushedFilters: [], ReadSchema: struct<id:int>

Does this PR introduce any user-facing change?

No.

How was this patch tested?

Unit test.

[wangyum]

@cloud-fan @sunchao

[cloud-fan]

This is tricky, as we need to be very careful about the reverse of date string parsing. Can we reference the detail of date string parsing in Cast and prove it's safe to do this optimization?

[beryllw]

UnwrapCastInBinaryComparison class is only responsible for processing casts related to numeric types, and it is dangerous to cast date types. Can we add the related conversion of partition columns to the PruneHiveTablePartitions rule?

[wangyum]

Close it, because this change may have potential data issue. Users can set spark.sql.legacy.typeCoercion.datetimeToString.enabled to true to restore the old behavior.