[InferenceClient] Third-party providers follow-up tasks

[Wauplin]

Originally posted by @Wauplin in #2757 (review)

As discussed offline, we'll have to take care about a few things:

• replace prepare_headers / prepare_payload / build_url by a unique prepare_request #2773
• revert providers from module-based to class-based (same as "hf-inference") #2778
• add documentation (examples with providers + maintain a provider <> tasks table) #2775
• Add text-to-video task and replicate's support for text-to-speech. #2784
• ASR parameters (+ likely T2I / TTS as well?) #2800
• implement proxy-ed calls (+ make sure we never leak HF token to another provider) #2780
• Switch to non-blocking APIs for image / video / audio generation.
• revamp VCR tests => server-side caching instead

All of this can be done in subsequent PRs. This PR is already big enough like this 😄

Thanks again for coordinating all this @hanouticelina ! It takes InferenceClient to a whole new dimension 🚀