[Fix] Support ndarray inputs in inpainting inferencer

[tandede]

Motivation

InpaintingInferencer declares np.ndarray as a supported input type, but its preprocessing pipeline always applies LoadImageFromFile and LoadMask. Arrays are therefore converted to strings and treated as file paths, which raises OSError: File name too long.

This fixes #1945.

Modification

• Build the image and mask loading stages independently according to each input type, so path/array combinations are all supported.
• Prepare array metadata consistently with file-loaded images, copy non-contiguous image views before tensor conversion, and normalize array masks with the same first-channel and binary semantics as LoadMask.
• Reject unsupported image and mask array shapes with clear errors.
• Add regression coverage for full AOT-GAN inference with ndarray inputs, mixed path/array inputs, negative-stride BGR views, mask normalization, source-array immutability, and invalid shapes.

BC-breaking

No. Existing path-based inputs keep the same loading pipeline and output.

Test plan

• pytest -q tests/test_apis/test_inferencers/test_inpainting_inferencer.py (9 passed)
• Relevant loading, masked-image, packing, and data-preprocessor tests (19 passed)
• Target source coverage: 100%
• Pre-commit checks on both changed files: passed

Checklist

• I have read and followed the contribution workflow.
• Pre-commit checks pass.
• The bug is covered by unit and end-to-end inference tests.
• The existing public input type already documents ndarray support; no separate documentation change is required.

[CLAassistant]

All committers have signed the CLA.