[ATOM-vLLM] Upgrade vLLM version to v0.22.0

[whx-sjtu]

Motivation

This PR is based on PR #750 and upgrades vLLM version to 0.22.0. This is necessary for further deepseek v4 enablement.

Technical Details

1. Adapt new is_ssm interface.
2. Adapt new MLA Prefill Backend.
3. Fix mamba_type interface change.
4. Fix topk_indices invalid value processing.
5. Add new dependences to Dockerfile.
6. Remove redundant self._use_cudnn_prefill

Test Plan

Full test on all ci cases.

Test Result

Case ID	Variant	Command Source	Flexible EM	Flexible Stderr	Strict EM	Strict Stderr
qwen3next_tp1	no-mtp	recipe	0.8226	0.0105	0.7354	0.0122
qwen3next_mtp_tp1	mtp	recipe	0.8173	0.0106	0.7316	0.0122
qwen35_397b_fp8_tp8	no-mtp	recipe	0.8468	0.0099	0.8316	0.0103
llama31_8b_tp1	no-mtp	recipe	0.7574	0.0118	0.6717	0.0129
glm5_tp8	no-mtp	recipe	0.9500	0.0060	0.9500	0.0060
glm47_tp4	no-mtp	recipe	0.9310	0.0070	0.9242	0.0073
glm47_mtp_tp4	mtp	recipe	0.9462	0.0062	0.9454	0.0063
deepseek_v32_tp4	no-mtp	ci-supplement	0.9522	0.0059	0.9515	0.0059
deepseek_v32_mtp_tp4	mtp	ci-supplement	0.9507	0.0060	0.9500	0.0060
dsr1_fp8_tp8	no-mtp	ci-supplement	0.9507	0.0060	0.9454	0.0063
dsr1_mxfp4_tp8	no-mtp	recipe	0.9393	0.0066	0.9371	0.0067
kimi_k25_tp4	no-mtp	recipe	0.9348	0.0068	0.9356	0.0068
kimi_k2_thinking_tp8	no-mtp	recipe	0.9348	0.0068	0.9333	0.0069

[zejunchen-zejun]

ATTN Refactor PR has been merged into main, so hexiang you can rebase the main