Fix: update app for mlx-swift-lm 3.x compatibility

[tmorales2000]

Fixes #77

The FastVLM app fails to build against current package versions due to breaking
changes in mlx-swift-lm 3.x and the migration of MLXLMCommon/MLXVLM out of
mlx-swift-examples into a new repository.

Changes

Package dependencies:

• Replace mlx-swift-examples with mlx-swift-lm 3.x
• Add swift-huggingface 0.9.x
• Update swift-transformers to 1.3.x (exposes Tokenizers as standalone product)

FastVLM.swift:

• Update attention mask type to ScaledDotProductAttentionMaskMode
• Fix createAttentionMask signature (windowSize parameter, single cache)
• Fix LoRAModel protocol conformance (loraLayers, loraDefaultKeys)
• Fix UserInput.Prompt handling (asMessages() removed in 3.x)
• Fix LMInput.ProcessedImage (imageGridThw -> frames)
• Fix ModelConfiguration loading API (url -> data in registry closures)
• Fix register() actor isolation (async, await)
• Qualify Tokenizer as MLXLMCommon.Tokenizer to resolve ambiguity

FastVLMModel.swift:

• Fix MLX.GPU.set(cacheLimit:) -> MLX.Memory.cacheLimit
• Fix loadContainer API (local directory + huggingFaceTokenizerLoader)
• Fix generate() closure (single Int token, accumulator pattern)
• Fix decode() label (tokens: -> tokenIds:)
• Fix GenerateCompletionInfo (output property removed)

Scripts:

• Replace wget with curl in get_models.sh and app/get_pretrained_mlx_model.sh

Testing

Tested on macOS with M1 Pro and M4 Pro:

• mlx-swift 0.25.6
• mlx-swift-lm 3.31.3
• swift-transformers 1.3.3
• swift-huggingface 0.9.0

App builds and runs successfully with 0.5B (FP16) and 1.5B (INT8) models.