call_stacks: rewrite for performance, add SQL implementation

[mwootton]

 - Pre-compute GPU times into dict instead of per-event LEFT JOIN
 - Bulk-insert all rows at once via executemany
 - Replace StackFrame class with mutable lists
 - Add pure SQL implementation (--sql) using window functions + recursive CTE
 - Improve CLI: add --sql flag, WAL mode, cache tuning, post-insert indexes
 - Both implementations produce identical output (verified on 2.6M rows)