[CI][Model] Stabilize MiniMax-M3-W8A8-A3 nightly validation (#15071)

## What this PR does

Unblocks the MiniMax-M3-W8A8-A3 nightly workflow so that the accuracy
benchmarks can run to completion.

This PR:

- switches the preliminary smoke test from the default completion
endpoint to chat completion with `max_tokens: 10` and `thinking_mode:
disabled`;
- disables EAGLE3 speculative decoding;
- disables the static kernel;
- reduces `--max-num-seqs` from 64 to 16 so the 64-request benchmark is
scheduled in stable waves;
- keeps GPQA `max_out_len: 65536` so long reasoning responses can
finish.

The existing accuracy baselines and thresholds are intentionally
unchanged. In particular, TextVQA remains at `baseline: 72` and
`threshold: 2`, and GPQA remains at `baseline: 92` and `threshold: 2`.

`enable_reduce_sample`, `FULL_DECODE_ONLY`, and the remaining W8A8
benchmark settings are also unchanged. The chat smoke arguments are not
passed to AISBench, so GPQA continues to run with its configured
reasoning behavior.

## Nightly failure analysis

[Nightly-A3 run
33080082724](https://github.com/vllm-project/vllm-ascend/actions/runs/33080082724)
checked out PR commit `30efb383da6b9d24bba0242dbbd9ae7279997258`. The
service became healthy, but the default `/v1/completions` smoke request
entered ACL graph replay and failed after the configured 60-second HCCL
timeout:

- `Communication_Error_Timeout(EI0002)`;
- AIV `AllReduce_group_name_5` Notify wait timeout;
- NPU error `507034`;
- EngineCore exited and the request returned HTTP 500.

No benchmark JSON was produced, so GPQA and TextVQA could not start. The
new `test_content: chat_completion` entry replaces that blocking request
path.

## Validation

Validated on a 16-card A3 host with the nightly CI image and the W8A8
model:

- service startup and `/health`: HTTP 200;
- chat-completion requests, including the thinking-disabled smoke
request: HTTP 200;
- 64 concurrent short chat-completion requests with `max_num_seqs=16`:
64/64 HTTP 200;
- `max_num_seqs=32` was rejected after a long-decode stress run
reproduced an ACL graph/HCCL timeout and engine exit;
- a 16K GPQA cap cannot meet the existing accuracy gate even if every
unfinished sample is counted as correct;
- a targeted 32K rerun recovered some long-output samples, but the
complete 32K run remained truncated, so the original 64K output limit is
retained.

The final [Nightly-A3 run
33146862269](https://github.com/vllm-project/vllm-ascend/actions/runs/33146862269)
confirmed that the workflow is no longer blocked:

- smoke request passed;
- GPQA completed all 198 samples with 89.90% accuracy;
- TextVQA completed with 72.93% accuracy and passed its original
threshold-2 gate.

GPQA remains slightly below its existing 90% lower bound. That accuracy
regression is present on the current main branch and is intentionally
left for a separate follow-up; this PR only restores a runnable,
non-blocking nightly path without weakening the accuracy gates.

- 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>