[ENHANCEMENT] Discard the intermediate recomputed outputs while restoring the original outputs

**Is your feature request related to a problem? Please describe.**

mcore has provided a [discard_output_checkpoint](https://github.com/NVIDIA/Megatron-LM/blob/core_v0.15.0rc5/megatron/core/tensor_parallel/random.py#L521C7-L521C30). It will do the recomputation via a backward hook. However, it does store the recomputation result and then restore the storage of the original outputs , followed by an inplace copy. See https://github.com/NVIDIA/Megatron-LM/blob/core_v0.15.0rc5/megatron/core/tensor_parallel/random.py#L587-L594. Note that the intermediate result can be released right after the copy to save memory aggressively. [However](https://github.com/NVIDIA/Megatron-LM/blob/core_v0.15.0rc5/megatron/core/tensor_parallel/random.py#L593) , mcore simply sets the intermediate tensors `None`, so that the memory management is returned to Pytorch/Python.

I'm wondering if we can mannualy release those tensors.

**Describe the solution you'd like**

Release the output within [the loop](https://github.com/NVIDIA/Megatron-LM/blob/core_v0.15.0rc5/megatron/core/tensor_parallel/random.py#L588-L590)
```
output_size = recomputation_output.untyped_storage().size()
output.untyped_storage().resize_(output_size)
output.untyped_storage().copy_(recomputation_output.untyped_storage())
recomputation_output.untyped_storage.resize_(0)
```

**Proposed implementation**
If you have a proposed implementation for the feature state it here or link to a PR.

**Additional context**
Add any other context or screenshots about the feature request here.