Delta ods hardening fixes

[ealexa05]

This PR addresses some issues with the new delta pipeline (added here: #203), and puts a new test table on the gamma instance (EDW.USE_TRANSACTION).

1. fix: Sparse updates no longer move rows to partition odin_year=0/odin_month=0
   The merge source derives odin_year/odin_month from the CDC batch's edw_inserted_dtm. When a sparse update batch carried no edw_inserted_dtm for a key, those degraded to 0 and were written verbatim. The row would silently move into the 0/0 partition (and out of partition-pruned query results) with internally inconsistent partition vs. data values. Partition columns now update via CASE WHEN source.edw_inserted_dtm IS NOT NULL THEN source.odin_year ELSE target.odin_year END, keeping the target's partition placement when the batch has no new timestamp.

2. fix: Null header__change_seq CDC records now raise instead of being silently skipped
   The watermark filter header__change_seq > ? never matches NULL in SQL, so null-seq CDC records were excluded before the existing null_count == 0 assertion could see them. The documented "CDC records carry a change sequence" invariant was thus unenforceable. The CDC read now includes an OR header__change_seq IS NULL arm with ORDER BY … NULLS FIRST, so any null-seq record surfaces at the head of the next batch and fails loudly.

3. fix: MERGE identifiers are quoted
   The MERGE key predicate and update/insert expressions used unquoted column names, so a column named like a SQL reserved word would break the merge. All column references in MERGE expressions are now quoted.

4. fix: History reads scoped to the current snapshot partition
   Schema discovery previously ran DESCRIBE over the table-wide glob with union_by_name, unioning columns across every historical snapshot. A column dropped upstream would reappear in silver as a phantom all-null column. All reads now use a snapshot=/**/*.parquet glob (DuckDB's ** matches zero or more directories, and hive partitioning still derives the snapshot column). Side benefit: the job no longer lists/reads footers for every historical snapshot's files on each run. The WHERE snapshot = ? filters remain as defense-in-depth.

[ealexa05]

Merging for velocity test on dev