Add MinimalAsyncEP + offset aware swiglu kernels (#3561)

[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

Squashed from PR #3561. Sync-free MoE token dispatcher (MinimalAsyncEP) +
offset-aware swiglu kernels, enabling full cudagraph for the MoE path.

Original: #3561