[graph_trainer] Replace stable_topological_sort with _move_overlap_nodes in fsdp_passes

[SherlockNoMad]

Stack from ghstack (oldest at bottom):

• [graph_trainer] Match Eager FSDP bucket order #3611
• [graph_trainer] Force cudagraph for MinimalAsyncEP on H100 + Run 4 results #3610
• Add MinimalAsyncEP + offset aware swiglu kernels (#3561) #3609
• [graph_trainer] Enable ChunkedCELoss for deepseek_v3 and qwen3 (#3248) #3608
• -> [graph_trainer] Replace stable_topological_sort with _move_overlap_nodes in fsdp_passes #3607
• [graph_trainer] Add DSv3 scaling run/profiling helper script #3606

Replace the two _stable_topological_sort calls in JointManualOverlapScheduler
with _move_overlap_nodes from upstream pytorch. This surgically moves only
the AG/RS chains instead of re-sorting the entire graph, reducing node
displacement from ~97% to ~5% and making the pass composable with downstream
graph passes.

• _manual_bucket_collectives: remove the no-op _stable_topological_sort(graph, {})
  call after bucketing — the graph is already valid after manual_bucket_collectives.
• _manual_reorder_graph: replace _stable_topological_sort(graph, overlap_deps)
  with _move_overlap_nodes(graph, overlap_deps, bucketed_node_types).

Depends on pytorch PR #184711.

Authored by Claude.

(cherry picked from commit 504dc71)