[Enhance] Accelerate training through cpu affinity on arm.

[Ginray]

Motivation

The current training script uses the torch.distrubuted function for multi-gpu training, which usually cannot achieve the best performance on arm cpu, and is more obvious on some models that require high data processing performance.

Modification

Provide cpu-friendly scripts, which are used in the same way as the original scripts

Use cases (Optional)

We verified on the 40+config of mmdetection, with an average improvement of about 10%

Checklist

1. Pre-commit or other linting tools are used to fix the potential lint issues.
2. The modification is covered by complete unit tests. If not, please add more unit test to ensure the correctness.
3. If the modification has potential influence on downstream projects, this PR should be tested with downstream projects, like MMDet or MMCls.
4. The documentation has been modified accordingly, like docstring or example tutorials.