Upgrade mlx-vlm and torchvision so Qwen3.5 multimodal will run

[perry2of5]

This PR upgrades libraries so --mllm will succeed. (Building for vision doesn't work on main right now.)

% git reset --hard origin/main
HEAD is now at b4fa030 Merge pull request #232 from sjswerdloff/fix/streaming-tool-call-content-leak
% rm -rf .venv uv.lock && uv venv --python 3.13 && source .venv/bin/activate && uv sync && uv pip install -e ".[dev,vision]"
Using CPython 3.13.11
Creating virtual environment at: .venv
Activate with: source .venv/bin/activate
Resolved 291 packages in 2.29s
error: Distribution `torchvision==0.18.0 @ registry+https://pypi.org/simple` can't be installed because it doesn't have a source distribution or wheel for the current platform

hint: You're using CPython 3.13 (`cp313`), but `torchvision` (v0.18.0) only has wheels with the following Python ABI tags: `cp310`, `cp311`, `cp312`

I tested by running the following:

vllm-mlx serve --served-model-name default mlx-community/Qwen3.5-35B-A3B-4bit

I ran vllm-mlx-chat in another window and it successfully responded to "say hi" and "tell me how many tables are in the attached ERD".

Originally there was a note that this depends on PR 215 or PR 243 being applied first, but PR 268 fixed the issue so that is no longer relevant

[perry2of5]

Updated to fix merge conflict. Works with

vllm-mlx serve --served-model-name default --mllm mlx-community/Qwen3.5-35B-A3B-4bit

and

vllm-mlx serve --served-model-name default mlx-community/Qwen3.5-35B-A3B-4bit

[waybarrios]

The torchvision bump makes sense, 0.18.0 has no cp313 wheels so this is needed for Python 3.13 users. One thing worth noting: torchvision 0.21.x requires torch 2.6.x, but the [vision] extras still declare torch>=2.3.0. pip will resolve it correctly through transitive deps, but bumping that floor to >=2.6.0 would make the constraint explicit. Up to you whether to include that in this PR or leave it for later.

[waybarrios]

Looks good to me so far.