v3: Add custom task register_pipeline

[Th3G33k]

Following feature request :

• Closes the ability to add new pipelines in transformers.js #659

Example use :

import {
  Pipeline,
  read_audio,
  register_pipeline,
  pipeline
} from '@xenova/transformers';

class AudioFeatureExtractionPipeline extends Pipeline {
  constructor(options) {
    super(options)
  }
  async _call(input, kwargs = {}) {
    input = await read_audio(input).then(input=>this.processor(input))
    let { audio_embeds } = await this.model(input)
    return audio_embeds
  }
}

register_pipeline('audio-feature-extraction', {
  pipeline: AudioFeatureExtractionPipeline,
  model: 'ClapAudioModelWithProjection',
  processor: 'AutoProcessor',
  default_model: 'Xenova/larger_clap_music_and_speech'
})

let pipe = await pipeline('audio-feature-extraction');
let out = await pipe('https://huggingface.co/datasets/Xenova/transformers.js-docs/resolve/main/jfk.wav')
console.log(out)

[Th3G33k]

We could also add custom pipelines directly to supported_task, if we removed the object freeze

const SUPPORTED_TASKS = Object.freeze({

[Th3G33k]

@xenova What do you think about this PR ? It would be similar to python transformers register_pipeline