feat: support internlm2

[zhyncs]

cc @Ying1123 @merrymercy @hnyls2002

tested with internlm2-chat-7b

python3 playground/reference_hf.py --model internlm2-chat-7b
python3 -m sglang.bench_latency --model internlm2-chat-7b --correct --output-len 16 --trust-remote-code

transformers

prefill logits tensor([291.2500, 293.7500, 364.5000,  ..., 307.5000, 332.2500, 257.0000],
       device='cuda:0')
prefill logits tensor([314.7500, 320.5000, 398.0000,  ..., 333.0000, 359.5000, 280.0000],
       device='cuda:0')
prefill logits tensor([332.2500, 334.2500, 404.2500,  ..., 351.5000, 374.5000, 293.0000],
       device='cuda:0')

sglang

prefill logits (final) tensor([[288.7500, 291.2500, 361.5000,  ..., 305.0000, 329.5000, 254.8750],
        [313.7500, 319.7500, 396.2500,  ..., 331.7500, 358.7500, 279.0000],
        [331.5000, 333.5000, 403.2500,  ..., 350.7500, 373.7500, 292.5000]],
       device='cuda:0', dtype=torch.float16)

[Ying1123]

On my machine (H100):
SRT:

prefill logits (final) tensor([[262., 262., 342.,  ..., 272., 290., 240.],
        [272., 276., 348.,  ..., 284., 302., 247.],
        [360., 362., 450.,  ..., 378., 392., 326.]], device='cuda:0',
       dtype=torch.bfloat16)

HF:

prefill logits tensor([265.5000, 264.5000, 345.0000,  ..., 274.7500, 293.0000, 242.7500],
       device='cuda:0')
prefill logits tensor([279.2500, 281.7500, 355.2500,  ..., 290.0000, 310.0000, 252.6250],
       device='cuda:0')
prefill logits tensor([359.7500, 361.2500, 448.5000,  ..., 377.0000, 390.5000, 324.5000],
       device='cuda:0')