[VLM] Avoid unnecessary dummy multimodal data during processing

[DarkLight1337]

Currently with the cache enabled, in BaseMultiModalProcessor._apply_hf_processor_mm_only we create both text and multimodal dummy data but the latter is unused, leading to unnecessary overhead for allocating NumPy arrays. This is especially devastating for models that support large image sizes such as Qwen2-VL series.

This PR optimizes this process by splitting up the dummy data generation into text and multimodal parts, so that we can create just the dummy text in BaseMultiModalProcessor._apply_hf_processor_mm_only.

Note: Multi-modal processors will be required to implement get_dummy_text and get_dummy_mm_data in a future release. This PR adds a warning asking the developer to do so if the processor hasn't done that yet.

Benchmark (#11196):

python benchmarks/mmmu_bench.py --model Qwen/Qwen2.5-VL-3B-Instruct --max-model-len 16384 --max-num-seqs 64 --num-prompts 250

Main branch

# Trial 1
Request throughput: 0.36 req/s
Total generated tokens: 1149
Token generation rate: 1.66 tok/s

# Trial 2
Request throughput: 0.51 req/s
Total generated tokens: 1151
Token generation rate: 2.33 tok/s

# Trial 3
Request throughput: 0.63 req/s
Total generated tokens: 1180
Token generation rate: 3.00 tok/s

This branch

# Trial 1
Request throughput: 0.54 req/s
Total generated tokens: 1173
Token generation rate: 2.51 tok/s

# Trial 2
Request throughput: 0.54 req/s
Total generated tokens: 1138
Token generation rate: 2.44 tok/s

# Trial 3
Request throughput: 0.57 req/s
Total generated tokens: 1191
Token generation rate: 2.74 tok/s

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[Isotr0py]

Nice, LGTM!