Thread image_grid_thw through Qwen3-VL, add smart-resize + multi-image

Per-image grid dimensions are now a real model input rather than being
inferred from sqrt(num_patches). This fixes correctness for non-square
images and lets one forward pass handle multiple images of different
sizes in the same prompt (concatenated patch sequence + per-image
grid_thw).

Featurizer changes:
- smart_resize mirrors qwen-vl-utils: preserves aspect ratio, rounds to
  a multiple of patch_size * merge_size, clamps total pixels between
  min_pixels and max_pixels
- :quality preset (:low / :medium / :high) over explicit pixel caps
- accepts a list of images of different sizes; concatenates patches,
  returns image_grid_thw of shape {num_images, 3}
- patches are emitted in windowed order so the merger reshape stays
  shape-agnostic (4 consecutive patches = one 2x2 merge block)

Vision encoder:
- image_grid_thw declared as a model input
- per-patch (row, col, grid_h, grid_w, image_id) derived from grid_thw
  via Nx ops; drives bilinear pos-embed interpolation and 2D rotary
- block-diagonal attention mask (image_id == image_id) so patches from
  one image cannot attend to patches from another
- patch merger and deepstack merger reshapes no longer assume a square
  single-image grid

Multimodal:
- image_grid_thw plumbed as an optional model input and forwarded into
  the vision sub-model

Validation:
- 13 new tests covering smart_resize aspect ratio, quality presets,
  multi-image concat, windowed-layout invariant, single+multi-image
  end-to-end with the tiny model
- Full fast suite: 285 passed, 0 regressions
- Real Qwen3-VL-2B-Instruct on a 640x480 COCO image: 8/10 top-token
  agreement vs HuggingFace transformers reference; top-3 identical and
  in same order

Refs #442.

Author: nyo16

lib/bumblebee/multimodal/qwen3_vl.ex

@@ -33,9 +33,15 @@ defmodule Bumblebee.Multimodal.Qwen3VL do
 
     * `"pixel_values"` - `{num_patches, flattened_patch_size}`
 
-      Pre-extracted image/video patches from the featurizer. The shape is
-      `{num_patches, channels * temporal_patch_size * patch_size * patch_size}`.
-      For a 384x384 image with default settings, this is `{576, 1536}`.
+      Concatenated, pre-extracted image/video patches from the featurizer.
+      Shape is `{num_patches, channels * temporal_patch_size * patch_size * patch_size}`.
+
+    * `"image_grid_thw"` - `{num_images, 3}`
+
+      Per-image grid dimensions `[temporal, height, width]` in patch
+      units. Threaded into the vision encoder so it can compute correct
+      per-patch positions for variable image sizes and multiple images
+      per prompt.
 
     * `"input_ids"` - `{batch_size, sequence_length}`
 
@@ -92,6 +98,7 @@ defmodule Bumblebee.Multimodal.Qwen3VL do
 
     %{
       "pixel_values" => Nx.template({num_patches, flattened_patch_size}, :f32),
+      "image_grid_thw" => Nx.template({1, 3}, :s64),
       "input_ids" => Nx.template({1, 1}, :u32)
     }
   end
@@ -114,7 +121,8 @@ defmodule Bumblebee.Multimodal.Qwen3VL do
       Bumblebee.build_model(spec.vision_spec)
       |> Bumblebee.Utils.Axon.prefix_names("vision_model.")
       |> Bumblebee.Utils.Axon.plug_inputs(%{
-        "pixel_values" => inputs["pixel_values"]
+        "pixel_values" => inputs["pixel_values"],
+        "image_grid_thw" => inputs["image_grid_thw"]
       })
 
     # Get vision embeddings using correct Axon.nx pattern
@@ -194,6 +202,7 @@ defmodule Bumblebee.Multimodal.Qwen3VL do
 
     Bumblebee.Utils.Model.inputs_to_map([
       Axon.input("pixel_values", optional: true, shape: vision_shape),
+      Axon.input("image_grid_thw", optional: true, shape: {nil, 3}),
       Axon.input("input_ids", shape: text_shape),
       Axon.input("attention_mask", optional: true, shape: text_shape),
       Axon.input("position_ids", optional: true, shape: text_shape),