[core] DiffusionPipeline.device deduction for split-device pipelines

[JingyaHuang]

Description

What does this PR do?

Follow-up to the discussion in #13718 (cc @sayakpaul @DN6), split out as requested.

• Issue

DiffusionPipeline.device returns the device of the first module component in alphabetical order of the __init__ signature (eg. text_encoder for Flux, Flux.2, Qwen-Image). When a pipeline is deliberately split across devices, DiffusionPipeline.device reports the device where locates the text encoder, for example:

# text_encoder stays on CPU
pipe.unet.to("cuda") 
pipe.vae.to("cuda")

pipe.device  # -> cpu, even though everything that computes is on cuda

This PR prefers any non-CPU, non-meta component, falling back to the previous first-component behaviour when there is nothing else to prefer. Backend-agnostic CUDA, XPU, and other accelerators, including TP-sharded models whose DTensor
parameters report their local shard device.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[sayakpaul]

But will we be able to run the pipeline with this configuration? For the test, I would also try to mimic the scenario in which this will be used in a minimal manner.