Parallelize llama4 implementation weight loading

[KWang1998]

Description

update code to support multi-threaded loading for llama4

Tests

Please describe how you tested this change, and include any instructions and/or
commands to reproduce.

Checklist

Before submitting this PR, please make sure:

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.

[github-actions]

Description

Start with a short description of what the PR does and how this is a change from
the past.

The rest of the description includes relevant details and context, examples:

• why is this change being made,
• the problem being solved and any relevant context,
• why this is a good solution,
• some information about the specific implementation,
• shortcomings of the solution and possible future improvements.

If the change fixes a bug or a Github issue, please include a link, e.g.,:
FIXES: b/123456
FIXES: #123456

Tests

Please describe how you tested this change, and include any instructions and/or
commands to reproduce.

Checklist

Before submitting this PR, please make sure:

• I have performed a self-review of my code.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have made or will make corresponding changes to any relevant documentation.

[jrplatin]

Can you include the time reduction / testing you performed (e.g. running MMLU / CI)?

[KWang1998]

Can you include the time reduction / testing you performed (e.g. running MMLU / CI)?

I tested locally with NEW_MODEL_DESIGN=True TPU_BACKEND_TYPE=jax python examples/offline_inference.py --task=generate --model=meta-llama/Llama-4-Scout-17B-16E-Instruct --max-model-len=1024 --tensor-parallel-size 8 --max-num-batched-tokens 1024 --max-num-seqs=1 --hf-config=meta-llama/Llama-4-Scout-17B-16E-Instruct --hf_overrides '{"architectures": ["Llama4ForCausalLM"]}'.
I found when using single thread, the time is 66.87 seconds.
When using multi-threads, the time is 84.07 seconds.