Feature: Add Qwen3-VL-8B-Instruct to the qwen_vl family

[jkzhang7]

Summary

Extend the qwen_vl family's existing Qwen/Qwen3-VL-2B-Instruct support to
Qwen/Qwen3-VL-8B-Instruct.

Why

Both checkpoints share the same architecture class
(Qwen3VLForConditionalGeneration, model_type: qwen3_vl) and only differ in
scale:

	2B (supported)	8B (this issue)
Text layers / hidden / heads	28 / 2048 / 16	36 / 4096 / 32
Vision depth / hidden	24 / 1024	27 / 1152

families/qwen_vl/config.py reads these dimensions from the checkpoint's own
config.json rather than hardcoding the 2B shape, so this is expected to be a
scale-up within the existing family rather than a new architecture.

Apache-2.0, ~19.7M downloads on Hugging Face.

Plan

• Add a manifest + test case for Qwen/Qwen3-VL-8B-Instruct under
  families/qwen_vl/tests/manifests/.
• Build and run the real GPU E2E (native generate / VL-inference CLI path)
  against the pinned checkpoint revision.
• If the 8B checkpoint needs tensor-parallel to fit comfortably, evaluate
  qwen_vl's existing TP builder path (already used for
  Qwen/Qwen2.5-VL-3B-Instruct).
• No shared-file changes expected (no new task type, no new CLI surface) —
  should stay entirely inside families/qwen_vl/.

Non-goals

Quantized precisions, additional Qwen3-VL sizes beyond 8B, streaming
generation.

[jkzhang7]

I can get a PR for this.