Refactor model URL constants

[msluszniak]

Description

Same as #382, but now I have no permission to push directly into separate branch on repo so changing to PR from fork.

[msluszniak]

Add info to change VERSION_TAG in model url constants. I created a https://github.com/software-mansion/react-native-executorch/pull/382which have introduced such tag. @pweglik in one #397 added models that are introduced with version tag for currently developed version aka 0.5.0. As a consequence, I probably introduce something like NEXT_DEV_VERSION_TAG in rebasing the PR. These will need to be changed once published a new version. I can add this one separately once rebase and push my PR.

This is what I meant to change the version tag. Right now, we have:

const VERSION_TAG = 'resolve/v0.4.0';
const NEXT_DEV_VERSION_TAG = 'resolve/v0.5.0';

And every time new version is release, we need to change the version of VERSION_TAG and NEXT_DEV_VERSION_TAG one up, and replace every NEXT_DEV_VERSION_TAG in the urls with VERSION_TAG since it is no longer next dev.
cc: @mkopcins @pweglik

[pweglik]

This is good suggestion, we should probably complete it alongside with : #359

[msluszniak]

@pweglik I don't know the list of models that require such presets from #359 . If you could send me some instructions I can implement this as e.g. present for LLM will have model, tokenizer, and config but text embeddings will have model, tokenizer, and mean pooling parameters. Otherwise, feel free to contribute to this PR.

[mkopcins]

You can try to match it by names, that should be fairly straight forward, for speech-to-text it is already defined, so that's that. You can add mean pooling falg alongside that. Regarding how to do that I think a simple json for all models on our huggingface as a starting point would be just fine. I am not sure how to generalize it beyond our models.

[msluszniak]

Currently, the whole codebase must be revised because there are situations where there is the following ussage:

await ImageSegmentationModule.load(DEEPLAB_V3_RESNET50);

and this one:

  const model = useImageSegmentation({
    modelSource: DEEPLAB_V3_RESNET50,
  });

Also onDownloadProgressCallback is used is a way that we currently do not support in this issue:

await LLMModule.load({
  modelSource: LLAMA3_2_1B_QLORA,
  tokenizerSource: LLAMA3_2_TOKENIZER,
  tokenizerConfigSource: LLAMA3_2_TOKENIZER_CONFIG,
  onDownloadProgressCallback: printDownloadProgress,
});

Ideally, we need to change every single usage in the codebase, to use presets, that are exposed to the end users.

[msluszniak]

@NorbertKlockiewicz do you have any idea how to deal with this?

[pweglik]

Regarding this:

Also onDownloadProgressCallback is used is a way that we currently do not support in this issue:

await LLMModule.load({
 modelSource: LLAMA3_2_1B_QLORA,
 tokenizerSource: LLAMA3_2_TOKENIZER,
 tokenizerConfigSource: LLAMA3_2_TOKENIZER_CONFIG,
 onDownloadProgressCallback: printDownloadProgress,
});

I think we can simply change it to:

await LLMModule.load({
  ...LLAMA3_2_1B,
  onDownloadProgressCallback: printDownloadProgress,
});

Regarding the other issue - yeah, I think we need to go through every place in docs and example apps and change it