Adding @huggingface/transformers brings unnecessary large dependencies

[samanhappy]

Adding @huggingface/transformers as a direct dependency pulls in onnxruntime-node (native binary, ~100MB+), onnxruntime-web, and sharp as non-optional dependencies. This significantly increases the Docker image size and install time, even though the ONNX runtime is never used (only the tokenizer is needed).

Consider using @huggingface/jinja + a lighter tokenizer-only package, or using tokenizers (the Rust-backed package) instead. Alternatively, since only BAAI/bge-m3 tokenization is needed, you could download and bundle just the tokenizer.json file and use a minimal SentencePiece/BPE decoder, avoiding the entire transformers dependency tree.

Originally posted by @Copilot in #672 (comment)

[samanhappy]

@josepcrespo I think we can try to reduce the image size if possible, as mentioned in the early review comment by Copilot. What do you think?

[samanhappy]

The image size comparison can be seen from: https://hub.docker.com/r/samanhappy/mcphub/tags

[josepcrespo]

Sure @samanhappy, we should try to implement the improvement suggested by Copilot. Would you like to start working on this yourself or assign me to do it when I can?

[samanhappy]

@josepcrespo It would be best if you could complete it, since you’re more familiar with the implementation from the previous PR. Of course, if you’re busy, I can also work on it.

[josepcrespo]

Yes, that makes perfect sense @samanhappy. I'll start working on it as soon as I have time.