[CPU] Bind threads and numa node for each TP rank

[chunyuan-w]

Motivation

This PR binds threads and numa node for each TP rank when device is CPU, which will improvement the performance on CPU.
For example, when running --tp 6 on a machine with 6 sub-numa nodes, with this PR, rank 0 will be bound to core 0-39 on node 0, rank 1 to core 40-79 on node 1, etc.

python3 -m sglang.bench_one_batch --batch-size 1 --input 1024 --output 8 --model  deepseek-ai/DeepSeek-R1  --trust-remote-code --device cpu --tp 6

NUMA:                    
  NUMA node(s):          6
  NUMA node0 CPU(s):     0-39,240-279
  NUMA node1 CPU(s):     40-79,280-319
  NUMA node2 CPU(s):     80-119,320-359
  NUMA node3 CPU(s):     120-159,360-399
  NUMA node4 CPU(s):     160-199,400-439
  NUMA node5 CPU(s):     200-239,440-479

Modifications

• A new OP init_cpu_threads_env is added in sgl-kernel/csrc/cpu/numa_utils.cpp to bind threads and numa node. UT for this OP is added in test/srt/cpu/test_binding.py.
• sgl-kernel/csrc/cpu/CMakeLists.txt is updated to link the numa lib. conda install -y libnuma numactl is needed as a pre-requisite, which will be added in the Dockerfile for Xeon ([CPU] enable CI for PRs, add Dockerfile and auto build task #6458).
• We don't need users to build vLLM CPU wheel anymore and updated the message in python/pyproject.toml.
• In python/sglang/srt/model_executor/model_runner.py, we add the threads and numa node binding according to the tp_size. When SGLANG_CPU_OMP_THREADS_BIND is not set, we only support tp_size to be smaller than the number of sub-numa nodes on the current machine and we will bind each tp rank to one sub-numa node. We only bind to physical cores while logical cores are excluded. If this is not satisfied, an error message will be thrown.