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

[SBrandeis]

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

[SBrandeis]

Ping for information: @julien-c @osanseviero @Wauplin