[vla] feat: consolidate LingBot VLA v2 training integration

[haichengv12]

Change Summary

1. support lingbotvla_v2. Improves performance, especially when GPUs are connected over a slow interconnect (e.g. PCIe instead of NVLink).
2. The gains come from four techniques: concurrent execution on separate CUDA streams, reduced-precision collectives between GPUs, operator fusion, and overlapping communication with computation.

[nullnonenilNULL]

@claude review this pr