`Pipeline.to()` offload guards only cover `cuda`/`xpu`: Ascend NPU misses the sequential-offload conflict error and the model-offload warning

[li-lizhe]

Describe the bug

In src/diffusers/modular_pipelines/modular_pipeline.py, two guards in Pipeline.to() are written as device_type in ["cuda", "xpu"]:

• line ~2691 — the sequential offload conflict check (raises the "you activated sequential model offloading … but are now attempting to move the pipeline to device" error)
• line ~2704 — the model offload warning (the operation succeeds but the offload benefit is silently lost)

Ascend NPU is also an accelerator where explicit device placement conflicts with offloading, but device_type == "npu" matches neither guard, so NPU users get silently broken offload behavior instead of the same clear error/warning CUDA and XPU users get.

device_type	Current guard	Expected
"cuda"	fires	fires
"xpu"	fires	fires
"npu"	silently skipped	should fire
"cpu"	skipped	skipped

Reproduction

import torch, torch_npu  # Ascend 910B
pipe = SomePipeline.from_pretrained(...)
pipe.enable_sequential_cpu_offload()
pipe.to("npu")   # no error raised, although the pipeline is sequentially offloaded

Environment

• diffusers main
• Ascend 910B NPU, torch 2.14 + torch_npu
• device_type resolves to "npu" for torch.device("npu")

Fix

Extend both guard lists to ["cuda", "xpu", "npu"]. PR: #14786