Add chunked prefill and limit mm per prompt options

[charitarthchugh]

The goal of this PR is to supplement the defaults to increase inference speed. I have tested this on a L40S and this has been working fine, with more than 2 weeks of combined uptime running pdf conversion.

Changes proposed in this pull request:

• --limit-mm-per-prompt allows VLLM to not allocate space on the GPU for the video encoder, thus increasing space for KV cache
• --enable-chunked-prefill basically streams the input tokens allowing for more efficient memory use.

Before submitting

• I've read and followed all steps in the Making a pull request
  section of the CONTRIBUTING docs.
• If this PR fixes a bug, I've added a test that will fail without my fix.
• If this PR adds a new feature, I've added tests that sufficiently cover my new functionality.

[jakep-allenai]

Sorry, been sick, just getting to this now. Let me run it locally and see what the numbers are looking like.

[jakep-allenai]

--limit-mm-per-prompt sounds like a great idea, but it seems like VLLM V1 is going to have chunked prefill always on by default. I'm waiting for benchmarks to run on our cluster here

[jakep-allenai]

Ok nice find on disabling the video encoder, really makes a difference on smaller GPU cards! There's a small syntax error, but I'm going to merge this in, as I want to push a PR today that moves things over to VLLM 0.11 officially as well.