[NPU]Add ZeRO-Infinity feature for NPU (#4809)
Add ZeRO-Infinity feature for NPU devices.
I add a new `async_io.py` in `op_builder/npu` and compilation
preprocessing judgment in `deepspeed_aio_thread.cpp` specifically for
NPU, which will be isolated from other devices such as the GPU and will
not affect each other.
See what we have already done in
https://github.com/microsoft/DeepSpeed/issues/4567 .
---------
Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>
