Optimize bulk operations by eliminating redundant LINQ operations

[oclockvn]

This commit addresses two performance bottlenecks in bulk operations:

1. Cache identity columns in CachedEntityMetadata

   • Previously GetIdentityColumns() filtered columns on every call
   • Now computed once during metadata construction like PrimaryKeyColumns
   • Eliminates repeated LINQ Where().ToList() operations
   • Estimated ~5-10% improvement for upsert with IdentityOutput=true

2. Use pre-filtered column lists in BuildMergeSql

   • Previously filtered columns to exclude identity on every upsert
   • Now passes both allColumns and columnsWithoutIdentity from cache
   • Eliminates redundant Where(c => !c.IsIdentity).ToList() at line 441
   • Estimated ~2-3% improvement for all upsert operations

Files modified:

• CachedEntityMetadata.cs: Added IdentityColumns property
• EntityMetadataHelper.cs: GetIdentityColumns() now returns cached value
• DbContextBulkExtension.cs: BulkUpsertAsync uses pre-filtered columns

[oclockvn]

@claude review this PR, response with agree or disagree and explain the reason