[BUG]Load DCP OOM

**Describe the bug**
when loading big enough model(20B PP4) from dcp format will cause OOM error
I have measure memory usage using `torch.cuda.memory._dump_snapshot(filename) ` and analyze snapshot file
show that
```
Megatron-LM/megatron/core/transformer/mlp.py:344 alloc 12344.00M
Megatron-LM/megatron/core/optimizer/distrib_optimizer.py:773 alloc19128.37MB
```
cost 31GB memory,it's unnecessary

I see @mikolajblaz have fix here,but it doesn't work,`mlp.py` will use lots of memory but no oom, until oom on optimizer states resume

on `distrib_optimizer.py`,state_dict should place on cpu,until `self.optimizer.load_state_dict`, this is also unnecessary

**To Reproduce**
using GPT model,using PP4 DP2 EP1 TP1 training a 20B model, each rank will hold 5B model Parameters,save to DCP format and resume this format

**Expected behavior**
resume 5B model should easy for memory

**Stack trace/logs**
If applicable, add the stack trace or logs from the time of the error.

**Environment (please complete the following information):**
- Megatron-LM commit ID: 2f1027d222e23b1ac5d917802bdddd01dfd10598
- PyTorch version:torch==2.6.0+cu126
- CUDA version 12.6
- NCCL version

**Proposed fix**
I will send a PR to fix it, where should I complete some testcase for that?

**Additional context**
Add any other context about the problem here.