Fix VLM CI and eliminate incorrect dependency on pixel values.

[lifuhuang]

Motivation

When I was merging change to support Kimi-VL, I noticed my CI was failing. Upon further investigation, I noticed some seemingly unexpected behaviors:

• the current VLMInputTestBase.test_understands_image is actually unexpectedly triggering the precomputed feature path, while it should have triggered the raw image path.
• Gemma3 and Qwen precomputed feature path are both taking dependency on pixel_values, which, based on my limited understanding, defeats the purpose of precomputed feature, which should have only taken dependency on precomputed feature themselves.

Modifications

1. Remove seemingly incorrect dependency on pixel values when precomputed features are provided.
2. Pass raw image url instead of processor output in test_understands_image tests.
3. Replace incorrect usage of processor with tokenizer in test_understands_image as per offline discussion with @mickqian

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[lifuhuang]

Unrelated to this change: improving naming from my last PR, sorry :P

[lifuhuang]

Discussed with @JustinTong0323, my understanding was wrong.