[PyTorch] Improve CP P2P efficiency

[yenchenlin]

Description

The original implementation saves each attention layer's whole K, V. Instead, we can discard used K, V to ensure each GPU only holds the current and the next K, V.

Memory profiling results using the toy unit-test

Original:

This MR:

One can see that the unused K, V are immediately released.

Type of change

• [ ] Documentation change (change only to the documentation, either a fix or a new content)
• [ ] Bug fix (non-breaking change which fixes an issue)
• [ ] New feature (non-breaking change which adds functionality)
• [ ] Breaking change (fix or feature that would cause existing functionality to not work as expected)
• [ ] Infra/Build change
• [x] Code refractor