[Bugfix, New Features] Ensure penalty sampler to be the first one in mixed samplers, accelerate TikToken, introduce pd disaggregation and separate acceleration on CPU backend

[huangzhengxiang]

1. Ensure penalty sampler to be the first one in mixed samplers. Penalty shall apply to the original logits.
2. Accelerate TikToken, leveraging Trie data structure.

prompt length	original O(n^3)	Trie O(n)
868	1.4s	1ms
24429	1h	34ms

(data on Snapdragon 8 Gen 3)

3. Introduce pd disaggregation and separate acceleration on CPU backend.

• Prefill and decode now can have separate configurations.
• CPU prefill becomes 15%～100% faster across 9 devices.

To name a few:

device	SoC	original	new	speedup
Huawei Mate 40 Pro	Kirin 9000	102	117	15%
Huawei Magic 6	Snapdragon 8Gen3	148	285	92%
Xiaomi 15 Pro	Snapdragon 8 Elite	258	368	42%

(speed unit: token/s, tested on Qwen2.5-1.5B int4 quant, block=128)