[CI][Model] Stabilize MiniMax-M3 nightly validation (#14930)

### What this PR does / why we need it?

This PR stabilizes the MiniMax-M3 BF16 A3 nightly configuration:

- explicitly uses `test_content: [chat_completion]` so the smoke request
exercises the model's chat endpoint instead of the default completions
endpoint;
- disables `enable_static_kernel`;
- disables `enable_reduce_sample`;
- pins `top_p: 0.95` for both GSM8K and TextVQA accuracy cases;
- tightens both accuracy thresholds from `5` to `2`.

These settings match the real-weight A3 validation path used for
MiniMax-M3. This PR only changes the nightly test case and does **not**
revert #13838.

### Does this PR introduce _any_ user-facing change?

No. This only changes nightly CI configuration and acceptance
thresholds.

### How was this patch tested?

Configuration checks:

- YAML parsing and an explicit assertion that `test_content` resolves to
`chat_completion`;
- explicit assertions for both accuracy cases, the performance
threshold, and the two additional-config flags;
- `git diff --check`.

Real-weight validation on 16 Ascend A3 logical devices with DP=2, TP=8,
EP enabled, and `FULL_DECODE_ONLY` ACLGraph:

| Case | Result | Target |
| --- | ---: | ---: |
| Service startup and chat smoke | HTTP 200, non-empty output | Pass |
| GSM8K accuracy | 96.875% (32 available samples) | 95 +/- 2: Pass |
| TextVQA accuracy | 71.2109375% (512 available samples) | 72 +/- 2:
Pass |
| Output token throughput | 17.5082 token/s | >= 20.2051 token/s: Fail |

The performance threshold is unchanged by this PR. The validation
container used a temporary uncommitted revert of #13838 while
investigating a separate configuration regression; per request, that
revert is not included here.

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

Signed-off-by: CXY-Katrina <katrina.cxy@gmail.com>