[Feat] Add native Tensor Parallelism support for HF backend

[YangKai0616]

Using TP to accelerate lm_eval for HF.

Usage:
With torchrun:

torchrun --nproc-per-node=4 -m lm_eval \
    --model hf \
    --model_args pretrained=google/gemma-4-31B-it,max_length=4096,add_bos_token=True,tp_plan=auto \
    --tasks gsm8k \
    --num_fewshot 5 \
    --batch_size 16 \
    --output_path ./GSM8K/31B

Or with accelerate launch:

accelerate launch --num_processes=4 -m lm_eval \
    --model hf \
    --model_args pretrained=google/gemma-4-31B-it,max_length=4096,add_bos_token=True,tp_plan=auto \
    --tasks gsm8k \
    --num_fewshot 5 \
    --batch_size 16 \
    --output_path ./GSM8K/31B

[CLAassistant]

All committers have signed the CLA.

[YangKai0616]

Hi @baberabb ,please help review this PR, thanks!
Cc @sywangyi

[YangKai0616]

Hi @baberabb , gentle ping on this PR. Let me know if you need any changes or further information from my side. Thanks!