Save CUDA Graph memory by reusing input and output tensors

[buptzyb]

Description

TE allows for passing multiple callables into make_graphed_callables() to share cudagraph pool to save memory. However, each cudagraph has its own input and output data buffer. This causes large memory overhead especially when _order is specified so that we must capture graphs for all microbatches. This PR allows for reusing the input or output tensors among cuda graphs captured in one make_graphed_callables() according to the pattern of _order. Note that reusing output tensors will cause extra DtoD memory copy in cudagraph execution.

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 refractor

Changes

Please list the changes introduced in this PR:

• Add reuse_graph_inputs option to make_graphed_callables() API.
• Add reuse_graph_outputs option to make_graphed_callables() API.

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