[main][feature][under updating]adapt for offload activation

[GeYuhong]

Description

This pr is used to adapt for offload activation (a new feature in Megatron-LM, NVIDIA/Megatron-LM#1752).

Offload activation select inputs of specific modules (such as core_attn, qkv_linear, router_fc1), offloading them to CPU in forward pass and reloading them to GPU in backward pass.

When offloading the modules that include weights (nn.Parameter), attributes of these weights (such as main_grad, grad_added_to_main_grad) are ripped off by torch. Therefore, this feature needs to modify the basic modules in TE (such as grouped_linear.py, 'layernorm_linear.py') to preserve these necessary attributes.

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
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Based on whether the input tensor contains the offloading_activation attribute, add support for retrieving the offload_activation flag in grouped_linear.py, linear.py, and layernorm_linear.py.
• Save the grad_added_to_main_grad attribute in forward pass and get it in backward pass in grouped_linear.py, linear.py, and layernorm_linear.py.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes