[WIP] [tests] test encode_prompt() in isolation

[sayakpaul]

What does this PR do?

encode_prompt() is often used in isolation to lift off memory requirements and to also compute prompt embeddings in many trainers. It's important that we test encode_prompt() properly.

This PR adds a test suite to do that. Since we cannot reliably map the outputs of encode_prompt() into keyword arguments of a pipeline call, I propose to add a test class attribute called prompt_embed_kwargs to pack the prompt embeddings related kwargs appropriately.

This PR is a PoC of what the changes might look like at the test level.

Once we agree on the common structure I will propagate to the rest of the pipelines.

The closest test we have currently is this:

diffusers/tests/pipelines/test_pipelines_common.py

Line 2165 in 15d4569

class SDXLOptionalComponentsTesterMixin:

But, IMO, this should be extended to the other pipelines on a more general basis.

TODOs

• Propagate to the rest of the pipelines once we agree on the structure.
• Fix encode_prompt() to work in isolation in the pipelines if needed.
• Remove SDXLOptionalComponentsTesterMixin if needed.

[sayakpaul]

@hlky now that the PRs have been merged to support this, let's brainstorm if this is how we want to test encode_prompt()?

[sayakpaul]

@DN6 a gentle ping.