[Feat] Dynamic Batch Feature (#3490)

[RFC](https://github.com/vllm-project/vllm-ascend/issues/3328) for more
details.
Add dynamic batch feature in chunked prefilling strategy, the token
budget can be refined to achieve better effective throughput and TPOT.

!!! NOTE: only 910B3 is supported till now, we are working on further
improvements.
Additional file for lookup table is required.

- vLLM version: v0.11.0rc3
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.0

---------

Signed-off-by: Cheng Wang <wangchengkyrie@outlook.com>

pyproject.toml

@@ -10,6 +10,8 @@ requires = [
     "pybind11",
     "pyyaml",
     "scipy",
+    "pandas",
+    "pandas-stubs",
     "setuptools>=64",
     "setuptools-scm>=8",
     "torch-npu==2.7.1.dev20250724",