[CI][Model] Stabilize MiniMax-M3 nightly smoke test (#15028)

## Summary

Builds on #14930 and stabilizes the MiniMax-M3 BF16 nightly case in two
ways:

- makes the chat-completions smoke request deterministic by setting
`max_tokens: 10` and disabling thinking;
- recalibrates the performance baseline from `20.83` to `18.13` token/s
using five repeated BF16 nightly runs with the current configuration.

The generic smoke request uses `max_tokens=10`. With MiniMax-M3 adaptive
thinking and the `minimax_m3` reasoning parser, all 10 tokens can be
returned as reasoning, leaving `message.content` null and failing the
smoke assertion. Setting `chat_template_kwargs.thinking_mode: disabled`
ensures a non-empty content response.

## Performance baseline

The performance reduction relative to the old baseline is expected for
the current MiniMax-M3 BF16 validation configuration introduced by
#14930. Five repeated nightly runs produced stable Output Token
Throughput results:

| Run | Output Token Throughput |
| --- | ---: |
|
[1](https://github.com/vllm-project/vllm-ascend/actions/runs/33035534563)
| 17.9823 token/s |
|
[2](https://github.com/vllm-project/vllm-ascend/actions/runs/33071251392)
| 18.1860 token/s |
|
[3](https://github.com/vllm-project/vllm-ascend/actions/runs/33080115364)
| 18.1831 token/s |
|
[4](https://github.com/vllm-project/vllm-ascend/actions/runs/33133138680)
| 18.1032 token/s |
|
[5](https://github.com/vllm-project/vllm-ascend/actions/runs/33136898007)
| 18.2097 token/s |

Statistics:

- mean: 18.1329 token/s
- median: 18.1831 token/s
- minimum: 17.9823 token/s
- maximum: 18.2097 token/s
- range: 0.2274 token/s
- population standard deviation: 0.0834 token/s

The new baseline is rounded from the five-run mean to `18.13` token/s.
The existing `threshold: 0.97` is retained, giving an effective pass
floor of `17.5861` token/s. This accommodates the observed run-to-run
variance while continuing to detect regressions greater than 3% from the
recalibrated baseline.

## Validation

Across the five dedicated BF16 nightly runs:

- chat smoke passed with HTTP 200, non-empty `content`, and
`reasoning=null`;
- GSM8K accuracy passed;
- TextVQA accuracy passed;
- performance results were stable and failed only against the obsolete
`20.83` token/s baseline.

- vLLM version: v0.27.1
- vLLM main:
https://github.com/vllm-project/vllm/commit/ba07e4a48fc951300d97eb506217dd530583dea3

---------

Signed-off-by: Ye Chenglin <1084697284@qq.com>
Signed-off-by: Ye Chenglin <2023200642@mail.nwpu.edu.cn>