[PyTorch] Add save_original_input in Linear/GroupedLinear to save memory

[hxbai]

Description

Add save_original_input in Linear/GroupedLinear. If it is on, the modules save the original inputs and only do rowwise quant in forward and colwise quant in backward (for blockwise and mxfp8 recipe). With this, in some scenarios, we can save memory usage at the expense of separated quantizations.

In the current implementation, the Linear/GroupedLinear always saves the cast input tensors for backward.
But this may increase the memory usage if this input tensor is used by multiple modules, such as:

• The output tensor of fused attn is saved by its self and used by the following proj_linear. We want the proj_linear to save the original input to save memory usage.
• For multi-latent attention in DeepSeek models, the output of input_layernorm is used by q_down_linear and kv_down_linear. We want the two linears to save the original input.

Besides, in Megatron-LM, we added an output-discarding recomputation that trickily releases some saved tensors and restores them by recomputation. This relies on that the saved tensor is the original output tensor.

Fixes # (issue)

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:

• transformer_engine/pytorch/module/linear.py
• transformer_engine/pytorch/module/grouped_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