SNOW-3489991: block flattening of dropped NEW columns in connect mode

In can_select_statement_be_flattened, the existing guard only refused to
flatten when a dropped NEW column was also is_referenced_by_same_level_column.
This missed the case where a future filter/sort (not yet added at select()
time) references the dropped column, generating invalid SQL.

In connect mode, extend the guard to block flattening unconditionally when
a NEW column is dropped. This preserves the defining subquery and keeps the
column visible to subsequent WHERE/ORDER BY clauses. Non-connect behavior
is unchanged.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: sfc-gh-joshi

src/snowflake/snowpark/_internal/analyzer/select_statement.py

@@ -2219,7 +2219,12 @@ def can_select_statement_be_flattened(
             state.change_state == ColumnChangeState.DROPPED
             and (subquery_state := subquery_column_states.get(col))
             and subquery_state.change_state == ColumnChangeState.NEW
-            and subquery_state.is_referenced_by_same_level_column
+            and (
+                subquery_state.is_referenced_by_same_level_column
+                # In connect mode, preserve the subquery so that future
+                # filter/sort clauses can still reference the dropped NEW column.
+                or context._is_snowpark_connect_compatible_mode
+            )
         ):
             return False
     return True