Fix attention_mask broadcasting for NPU compatibility

[chang-zhijie]

This PR resolves the unsupported atten_mask shape error when running attention with NPU (Ascend) devices.

Problem:
The NPU’s fusion attention operator (e.g., npu_fusion_attention) does not support automatic broadcasting for attention masks.
When a mask of shape [batch, seq_len] or [batch, 1, 1, seq_len] is passed, the operator fails with an error similar to:
get unsupported atten_mask shape, the shape is [B, 1, 1, S] – while only shapes like [B, N, S, S], [B, 1, S, S], [1, 1, S, S], or [S, S] are accepted.

Solution:
When running on NPU (is_torch_npu_available()), explicitly expand the mask to [batch, num_heads, seq_len, seq_len] to satisfy the operator’s shape constraints.
For non‑NPU devices, the mask is kept in broadcastable form ([batch, 1, 1, seq_len]) for efficiency.

Reference:
Ascend NPU fusion attention API:
https://www.hiascend.com/document/detail/zh/Pytorch/730/apiref/torchnpuCustomsapi/docs/context/torch_npu-npu_fusion_attention.md

Tested:
✅ NPU (Ascend) with explicit expansion – no shape error
✅ GPU / CPU with broadcastable mask – no performance regression

[chang-zhijie]

@yiyixuxu This PR is a follow-up to #13432, adding NPU hardware support for ERNIE-Image. Could you please help review it? Thanks!