Use inference types from `@huggingface.js/tasks` inside `@huggingface/inference`

Now that we have some generated types for input / output of each task, it would be nice to start using them for typing the inference clients.

It will also help us identify incompatibilities / weak points inside the task JSON schema specs, if any, and iterate to correct them :)

Tightly coupled to #360 (although independant)

cc @coyotte508 @mishig25 