Qwen3VL: default per-image resolution to a 2,560 vision-token budget

[spokvulcan]

Fixes #396.

The qwen3_5 ViT runs global O(patches²) attention with no window structure, so an uncapped full-resolution image (several published configs ship a ~16 MP longest_edge that never clamps) can allocate a tens-of-GB attention score matrix. Measured on a 27B Qwen3.6 checkpoint: 47 GB peak / 121 s prefill for one full-screen screenshot; ~21 GB with this cap.

• Default each image — and each video frame — to a 2,560 vision-token budget: min(config.maxPixels, 2560 * factor²), mirroring the Qwen2.5-VL clamp.
• Honor an explicit processing.maxPixels override, so full resolution stays available as an opt-in.
• Only what the tower sees is downscaled; stored image bytes are untouched.
• The budget multiply is guarded with multipliedReportingOverflow so a pathological patch/merge size degrades to the config ceiling instead of trapping, and the image and video preprocess paths clamp through the same config.maxPixels / config.minPixels accessors so the two clamp sites cannot drift.

On the default value: if you'd rather keep upstream's current behavior as the default, I'm happy to reshape this as a pure opt-in knob — the part that matters for on-device use is that a per-image budget exists.

[spokvulcan]

Lint fixed.

[spokvulcan]

Rebased onto current main. The mac_build_and_test failure here was the CudaBuild package-plugin validation issue that #404 fixed on main (-skipPackagePluginValidation) — it predates this PR's code, which never got compiled in that run. Verified locally on the rebased branch with the CI steps (build-for-testing + MLXLMTests): builds clean, all tests pass. The branch's swift-format-603 style commit was dropped during rebase since #404 already contains the identical change.