fix: skip weights defined in create_weights for pp.

[yuxianq]

Since DecoderModel's __pp_init__ is called before DecoderModelForCausalLM's __post_init__, it fails to skip weights for those weights defined in create_weights, which is created inside __post_init__.
We call DecoderModel's __pp_init__ inside DecoderModelForCausalLM's __pp_init__ to fix it, since DecoderModelForCausalLM's __pp_init__ is called after its __post_init__.

[yuxianq]

/bot run --disable-fail-fast --add-multi-gpu-test

[tensorrt-cicd]

PR_Github #5739 [ run ] triggered by Bot

[Barry-Delaney]

LGTM. Local tests passed.

[amukkara]

does any CI test fail before this change?

[tensorrt-cicd]

PR_Github #5739 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4195 completed with status: 'FAILURE'

[yuxianq]

does any CI test fail before this change?

@amukkara No, @Barry-Delaney get OOM issue when running python examples/pytorch/quickstart_advanced.py --model_dir /llm-models/DeepSeek-R1/DeepSeek-R1-W4AFP8 --tp_size 2 --pp_size 2 --moe_ep_size 1 --moe_tp_size 2 on H200x4. After this PR, this test can pass. Our CI does not contain any DeepSeek-R1 test now.

[yuxianq]

/bot run --disable-fail-fast --add-multi-gpu-test

[tensorrt-cicd]

PR_Github #5812 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #5812 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4257 completed with status: 'FAILURE'

[yuxianq]

/bot run --disable-fail-fast --add-multi-gpu-test

[tensorrt-cicd]

PR_Github #5871 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #5871 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #4302 completed with status: 'SUCCESS'