SIGSEGV in AnalysisPredictor::PreparePirProgram on Linux ARM64 (Docker/Apple Silicon)

### bug描述 Describe the Bug Environment: * OS: Linux ARM64 (inside Docker on macOS Sequoia) * Hardware: Apple M1/M2/M3 (ARM64) * Versions: PaddlePaddle 3.0.0b2 through 3.2.2 * Python: 3.12-slim Description: A consistent segmentation fault occurs during PaddleOCR initialization when running on Linux ARM64 within a Docker container. The crash is triggered regardless of whether FLAGS_enable_pir_api=0 is set, suggesting the PIR path is being forced during predictor initialization. C++ Traceback: 1 0 paddle_infer::Predictor::Predictor(paddle::AnalysisConfig const&) 2 1 paddle::AnalysisPredictor::Init(...) 3 2 paddle::AnalysisPredictor::PreparePirProgram() 4 3 paddle::AnalysisPredictor::SaveOrLoadPirParameters(bool) 5 4 std::filesystem::__cxx11::path::~path() <-- CRASH HERE Steps to Reproduce: 1. Run a python:3.12-slim Docker container on an Apple Silicon Mac. 2. pip install paddlepaddle==3.2.2 paddleocr==3.4.1. 3. Initialize PaddleOCR(lang='en'). 4. Observe SIGSEGV during the first call to ocr(). Suggested Fix: The crash originates in paddle/fluid/inference/api/analysis_predictor.cc. It appears std::filesystem usage in SaveOrLoadPirParameters is not thread-safe or has memory alignment issues on aarch64 Linux builds. Bypassing the PIR executor (downgrading to 2.6.2) resolves the issue, but the 3.x line remains unusable on ARM64 Docker. ### 其他补充信息 Additional Supplementary Information _No response_