Add Qwen 3 VL Support (Dense) (#414)

* Add Qwen3-VL model support

* Clean up model implementation

* Removing debug logging

* Revert ContentView and VLMEvaluator

* Refactor Qwen3VLLanguage and Qwen3VLVision models

* Replace custom splitFeatures() with native MLX split(indices:)

Use MLX's built-in split() function instead of manual Swift slicing loop. Converts size array to cumulative indices for MLX's split(indices:) method. Removes 12 lines of unnecessary Swift code.

* Vectorize rotaryPositionEmbedding() using MLX broadcasting instead of nested loops

Replace 5-nested-loop coordinate generation with MLX vectorized operations:
- Changed from Swift Array loops (~113K iterations) to MLX broadcasting
- Use expandedDimensions() for Cartesian product generation
- Apply arithmetic on broadcasted tensors instead of element-wise appends
- Use stacked() and tiled() for final coordinate organization

This is a direct port of Python's meshgrid pattern already used in
positionalEmbeddings(). Expected speedup: 50-75% for this function.

* Refine rotaryPositionEmbedding() vectorization implementation

- Add guard statement to skip empty grids (mergedH or mergedW == 0)
- Use .reshaped() instead of multiple expandedDimensions() calls
- Move mergeScalar computation outside the loop for efficiency
- Explicit broadcast to target shape before flattening
- Cleaner variable naming and comments

This approach is more efficient and matches the Python implementation pattern.

* Refactor Qwen3VL and Qwen3VLVision for improved cumulative index calculations

* Refactor ContentView and VLMEvaluator

* Update model configuration in VLMEvaluator

* Update model configuration in VLMEvaluator to use `smolvlm`

* Remove Qwen3VL_PositionIDs, Qwen3VLConfiguration, Qwen3VLLanguage, and Qwen3VLVision files

Author: rudrankriyam

Applications/VLMEval/ContentView.swift

@@ -416,7 +416,6 @@ class VLMEvaluator {
 
                 let stream = try MLXLMCommon.generate(
                     input: lmInput, parameters: generateParameters, context: context)
-
                 // generate and output in batches
                 for await batch in stream._throttle(
                     for: updateInterval, reducing: Generation.collect)