Add Idefics3Model for ColSmol

[akshayballal95]

This PR adds the Idefics Model mainly for use with ColSmol, but the support can be extended to use as a CausalVLM for use with models like Idefics3-Llama.

Although the model works well as intended, and the output tensors are acceptably close to the Transformer library counterpart, the execution speed is quite low (about 75% slower than the transformer). Upon inspection, I found that one particular operation takes an inordinately long time. It's this particular operation:

https://github.com/akshayballal95/candle/blob/e71c90e8d1db5ed84ff65cef8cc4f9f3301ff30e/candle-transformers/src/models/idefics3/model.rs#L596-L599

  let special_image_token_indices = input_ids
      .eq(self.image_token_id as f64)?
      .nonzero()?
      .repeat((1, embed_dim))?;

Specifically, the nonzero operation contributes to the time. But I don't think the issue is with nonzero. I have used this operation in other parts of the code, and it runs quite fast everywhere else—just not in this particular case. I understand that this involves moving the tensors to the CPU, but when I keep input_ids on the CPU, this operation becomes faster. However, strangely, the inference time for Llama increases. Also, replacing nonzero with other operations like cumsum has the same problem.

Here are the times for major parts of the model:

TextEmbed time: 22.219µs
Vision Transformer time: 34.432157ms

Input Merging time: 466.175661ms
-> Special Image Token Indices time: 466.096411ms ❌

LLama Inference time: 15.215606ms

Let me know if more information is needed or if this is normal behavior.