perf: cut train-loop sync/retrace overheads in OneVision Encoder

[Luodian]

summary

• fix grad-accum step boundary + scheduler stepping alignment
• cut per-step sync points in training/train.py (remove hot-path .item() / host sync patterns)
• add safer compile control: --compile_backend {auto,none,inductor,aot_eager,eager}
• in auto, disable compile on mixed dali_type to avoid retrace storms
• cache RoPE frequency grids and skip dense-path identity gather in encoder forward
• reduce dataloader overhead in data_v2* wrappers (drop redundant .cuda(), raise prefetch/thread defaults)
• reduce per-step frame-sampling allocation pressure in residual branch

why

• less GPU<->CPU sync in hot path
• fewer dynamic graph recompiles under mixed input signatures
• less repeated tensor construction / memory traffic per step
• better input pipeline overlap for image branches

replay & exact-match notes

• This PR is performance-oriented and does not guarantee byte-exact identity against historical runs under default settings.
• For closest parity with previous behavior, run with:
  • --compile_backend none (disable torch.compile wrapper)
• If you need strict checkpoint-comparison parity too, keep all non-performance knobs unchanged (dataset list/order, random seeds, worker layout, and DALI environment/config).
• Note: if your environment previously used old data_v2.py / data_v2_ocr.py defaults for prefetch_queue_depth=1, this PR changes them to 3; this can change data ordering timing while keeping sample values unchanged.

validation

• python3 -m py_compile training/train.py onevision_encoder/modeling_onevision_encoder.py dataloader/data_v2.py dataloader/data_v2_ocr.py dataloader/data_v2_multi_res.py