ui: Three flamegraph perf wins: skip empty walk, pre-floor trim, parentCum

Three independent bit-identical optimizations that together shave ~20%
off native bottom-up pipeline time (measured -16s on a 17M-walk-row
heap-graph trace, 83s -> 67s). All three are surgical: same output,
no algorithmic change, just removing wasted work.

1. Skip the empty UNION side in the hash walk. For BOTTOM_UP the
   downwards walk has empty inits (showDownward=FALSE) so it produces
   no rows, but graph_scan still does setup work. For TOP_DOWN the
   upwards walk is similarly empty (isPivot is never true). In the TS
   layer, only call whichever macro actually produces rows.

2. Pre-floor the trim propagation. The floor threshold (min_value) is
   now applied at the scan edge filter instead of inside the
   per-node IIF. Dest nodes below min_value are unreachable from the
   seeded roots, which is exactly the "dead" semantic the original
   implementation achieved via +inf propagation. On this trace the
   scan input shrinks from 8.4M rows to ~108K rows; the propagation
   step drops from 12.65s to 1.03s. Output bit-identical.

3. Precompute parentCumulativeValue. resolve_groups already does
   LEFT JOIN $grouped p to assign parentId; pulling p.cumulativeValue
   through as a new column is free. trim_with_placeholder now carries
   it forward (placeholders use MIN(d.parentCumulativeValue) since all
   dropped children of a parent share it). global_layout reads it
   directly from s instead of doing its own LEFT JOIN $merged p.
   global_layout drops from 13.2s to 1.0s.

Also adds ORDER BY id to trim_with_placeholder. The merged UNION ALL
wasn't emitting rows in dense id order, which the journal from the
prior commit noted makes graph_scan 5x slower downstream in
global_layout. Fixing it recovers global_layout's speedup from E.