Fix: FLUX.2 encode_prompt dtype and add bitsandbytes int8+bfloat16 warning

[CoderSATTY]

What does this PR do?

Fixes #13772

Two related fixes for the Flux2Pipeline:

1. Added a defensive warning for bitsandbytes 8-bit + bfloat16 quantization.
   This specific combination causes precision loss (bfloat16 is internally downcasted to float16 during MatMul) which accumulates and corrupts images in FLUX models. The warning alerts users at pipeline initialization and suggests using NF4 4-bit quantization instead.

2. Fixed encode_prompt() for pre-computed embedding workflows.

   • Skips prompt string formatting if embeddings are already provided.
   • Automatically casts pre-computed embeddings to the exact precision (dtype) and device expected by the pipeline. This prevents the silent precision mismatches that were causing corrupted/noisy outputs when passing embeddings between different pipeline instances.

Testing

• Tested on NVIDIA A100-80GB via Modal.
• Verified that the pipeline correctly warns when loaded with BitsAndBytesConfig(load_in_8bit=True) and torch.bfloat16.
• Verified that single-pipeline and two-phase workflows function perfectly and without warning when using NF4 quantization.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case: Bad image output for Flux.2-dev, rocm, quantization and separate prompt encoding sequence #13772
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

Who can review?

@yiyixuxu @sayakpaul