Use inference types dataclasses everywhere

[Wauplin]

We now have specs for all inference types with the goal of unifying APIs across the HF ecosystem, both in Inference APIs and transformers/diffusers/etc. pipelines. Since #2036, we have auto-generated Python dataclasses meant to be reused as much as possible. At the moment the output types are used for most InferenceClient outputs. This issue is meant to list remaining things to address.

• use new types for text-generation output (requires specs for stream param)
• remove all pydantic from our codebase (not in use anymore)
• better auto-generated docstrings (especially to document arguments)
• generate input parameters for InferenceClient from input types
• generate InferenceClient docstrings from input types
• proper generation script in CI (open PRs automatically on update?)
• reuse in transformers pipelines
• reuse in diffusers pipelines
• ?

cc @SBrandeis @julien-c

Related:

• jsonschema specs: https://github.com/huggingface/huggingface.js/tree/main/packages/tasks/src/tasks
• generated Python classes: https://github.com/huggingface/huggingface_hub/tree/main/src/huggingface_hub/inference/_generated