Add support for Idefics 2

[EricLBuehler]

This PR adds support for our first multimodal model: Idefics 2 (https://huggingface.co/HuggingFaceM4/idefics2-8b)!

Implementation TODOs:

• [x] VisionTransformer
  • [x] Encoder
    • [x] Attention
    • [x] MLP
  • [x] VisionEmbeddings (pending issue 2185 or a Tensor::bucketize function)
• [x] Connector
  • [x] MLP
  • [x] PerceiverLayer
• [x] Model
  • [x] Forward pass
    • [x] Remove padding images
    • [x] Generate the patch/pixel attention mask (pending a Tensor::unfold function)
    • [x] Run vision submodel and connector submodel
      • [x] Allow Mistral to run trained embedding head on any input tokens
      • [x] Inputs merger to inject embeddings correctly
    • Pass input to Mistral model
      • [x] Allow Mistral to take an embeddings vector instead of using the trained embedding head.
• [x] Image processor analogous to Idefics2ImageProcessor
  • [x] Resizing
  • [x] Rescaling
  • [x] Normalization
  • [x] Padding
    • [x] Generate pixel attention mask for padded images
      • [x] Pass and use in input injection
  • [x] Create pixel values tensors
• [x] Vision Model Pipeline
  • [x] Add a VisionModel trait similar to NormalModel
  • [x] Add a ModelCategory: vision, text, embedding etc
  • [x] Handle sequence scheduling with image dimensions
  • [x] Abstract input preparation logic
    • [ ] Handle padding to same, resized shape, across batch dimension
• [x] Add proper handling of chat templates
  • [x] Load preprocessor/processor config JSON files
  • [x] Support configuration of inputs processor via preprocessor
• [x] Messages API generalization
  • [x] Support OpenAI compatible method of specifying images
  • [x] Update messages to optionally encode type (akin to examples here).
  • [x] Use processor config to abstract the chat template application process
• [x] HTTP API
  • [x] Handle decoding from base64
  • [x] Support loading from HTTP.
• [ ] Rust API
• [ ] Python API

Other TODOs:

• [x] Introduce model type enum to reject mixing of text/multimodal models in speculative decoding
  • Perhaps introduce VisionModel akin to NormalModel.
• [ ] Ergonomic API support (OpenAI compatible on the HTTP side, but hopefully nicer on the Rust/Python side)
• [ ] Support device mapping
• [ ] Support ISQ

Pending issues:

• huggingface/candle#2185
• mokeyish/candle-ext#7