Make Ascend NPU available (#3831)
* Make Ascend NPU available
NPU accelerator support is introduced in (#3595).
This commit provides two enhancements:
1. Add a new accelerator_name 'npu' for choosing, it can be specified
by environment variable or auto detected.
2. Optimize auto detect code in get_accelerator to avoid too many
layers of exception throwing.
* Use DS_ACCELERATOR_LIST for overriding accelerators
When detecting override accelerators there's an error message to show
all support accelerators, using an accelerator list instead of hard
coding accelerator names in this message.
And fix code format issue(yapf).
* Add HCCL backend
HCCL is the distribute backend of Ascend NPU, it already implemented in
npu plugin for pytorch (https://gitee.com/ascend/pytorch). Add HCCL
backend as a not implemented backend to avoid not supported warning.
* Add NPUNotImplementedBuilder
Ascend NPU does not implement any op yet, leave npu folder empty will
throw NoneType[op_name] when not supported op is called. Add this
NPUNotImplementedBuilder as the default builder.
* Optimize builder search logic
1. cpu and other backend implement their ops in sub dirs under
op_builder, cuda_accelerator should skip these sub dirs.
2. Each backend will have its own NotImplementedBuilder, add device
prefix to this class to distinguish.
* Change the unimplemented builder name to the same for each backend
