Preserve projection metadata and dialect-aware CREATE parsing

[osipovartem]

Summary

• delegate non-DataFusion CREATE statements back through the configured SQL dialect, restoring Snowflake CREATE ICEBERG TABLE, CREATE STAGE, and fork-provided CREATE EXTERNAL VOLUME parsing used by Rustice
• reserve DataFusion's CREATE EXTERNAL TABLE parser only for an actual [TEMPORARY] TABLE prefix
• preserve metadata-changing projections when projection pushdown rewrites unions, filters, cross joins, hash joins, and sort-merge joins
• preserve outer field metadata for direct-column projection chains while continuing to derive metadata for computed expressions such as Arrow extension-type casts
• align inherited unit, SQL logic, and TPCH expectations with the fork's Snowflake half-away-from-zero decimal semantics and named UDTF arguments
• fix the hash-join spill rustdoc link under RUSTDOCFLAGS=-D warnings
• keep grace hash-join spill partition routing independent from intentional force_hash_collisions, while preserving collision coverage inside each partition's hash table

This prevents DataFusion 55's physical ProjectionPushdown optimizer from restoring source PARQUET:field_id metadata that a logical projection intentionally removed, while avoiding the opposite regression where native extension metadata was lost.

DataFusion 50/53 port audit

The three Embucket commits on DF50 were carried into the DF53 fork lineage; both histories were checked before classifying the items below.

Retained on DF55:

• Snowflake SQL extensions, PIVOT/UNPIVOT lowering, named table-function arguments, and the planner APIs used by Rustice
• decimal division and AVG half-away-from-zero semantics
• hash-join spilling
• Embucket SQLParser features through the Rustice [patch.crates-io] pin, including EXTERNAL VOLUME, standalone BEGIN, SHOW VARIABLES, timestamp precision, partitioning, and Snowflake string escapes

Superseded or owned elsewhere:

• the old logical/protobuf Pivot node is replaced by DF55 lowering to standard aggregate/union plans
• the old subquery/CSE patches are superseded by DF55's redesigned decorrelation path
• timestamp TO_DATE coercion is upstream in DF55
• REGEXP_SUBSTR is implemented and tested in Rustice

Not blindly ported:

• the untested df53 boolean-vs-integer comparison coercion and permissive non-string || coercion
• public visibility changes for APIs no longer used by Rustice

Those coercions broaden upstream behavior and should only return behind Snowflake differential correctness tests.

Verification

• cargo +1.97.0 fmt --all -- --check
• cargo +1.97.0 test -p datafusion-sql (91 unit, 572 integration, 12 doctests passed; 1 doctest ignored)
• cargo +1.97.0 clippy -p datafusion-sql --all-targets -- -D warnings
• cargo +1.97.0 test -p datafusion-physical-expr --lib (1,586 passed; 2 ignored)
• RUSTDOCFLAGS='-D warnings' cargo +1.97.0 doc --document-private-items --no-deps --workspace
• all six previously failing collision-mode spill regressions pass locally, including all join types, NULL equality, recursive repartitioning, and CollectLeft
• cargo +1.97.0 clippy -p datafusion-common -p datafusion-physical-plan --lib --features datafusion-physical-plan/force_hash_collisions,datafusion-common/force_hash_collisions -- -D warnings
• focused decimal division/AVG, hash-join spill, PIVOT/UNPIVOT, named UDTF, projection metadata, and SQL logic regressions
• Rustice cargo +1.97.0 test -p executor --lib --locked (436 passed; 7 intentionally ignored), including fork-backed CREATE EXTERNAL VOLUME parsing

Required by Embucket/rustice#59.