[Test] Full-iteration CG + FSDP

[erhoo82]

Support full-iteration cuda graph for FSDP

• This needs a patch in TE 2.10
• A fix to trigger the param AG hook in the current forward path (not in the optim.step of the previous step)

[erhoo82]

@sanandaraj5597 to confirm on Megatron Bridge at NeMo 26.02 once all updates from Jack Chang is done.

[erhoo82]

• One TE fix: [PyTorch] Bunch of fixes for cpu offloading NVIDIA/TransformerEngine#2535
• Another TE fix to come
• One MCORE fix to come

[erhoo82]

This is assigned to Ritesh Patel of MLPerf team.
CC @ko3n1g

[erhoo82]

Pushing this to 26.04.