[Feature] MindNLP 0.6 performance optimize

[lvyufeng]

Motivation

Reduce the host-bound overhead of MindSpore opeators, optimize Dispatch machanism.

Dispatch

• Tensor.device: use mindspore native api instead of python added (Only MindSpore >= 2.7.0 supported)
• Dispatcher: use a triky method, only refer the first Tensor input, do not compare with each other with python(will speedup 2x when small models)
• Multi-backend: still support aclnn and aclop(old primitive ops)
• API patch: use a switch to open or close.

New environments

• ENABLE_DISPATCH: whether use python dispatcher
• ENABLE_PYBOOST: whether use pyboost ops(used for OrangePi and GE graph)
• ENABLE_API_PATCH: speedup models