Add TorchScript/JIT module support for model inference

[tphakala]

Summary

• Add new jit package with Module type for loading and running TorchScript models
• Support single and multi-output forward passes (Forward, ForwardMulti) for models like BirdNET v3.0
• Implement device targeting (CPU/CUDA), eval/train mode switching, and thread-safe cleanup
• Leverage Go 1.25 features: runtime.AddCleanup, iter.Seq2, testing/synctest, CGO escape hints

中文摘要 (AI翻译，如有错误请见谅)

新增 TorchScript/JIT 模块支持，实现模型推理功能：

• 新增 jit 包，包含用于加载和运行 TorchScript 模型的 Module 类型
• 支持单输出和多输出的前向传播（Forward、ForwardMulti），适用于 BirdNET v3.0 等返回多个输出的模型
• 实现设备选择（CPU/CUDA）、评估/训练模式切换，以及线程安全的资源清理
• 使用 Go 1.25 新特性：runtime.AddCleanup、iter.Seq2、testing/synctest、CGO 逃逸提示

新增文件：

• jit/jit.go - 高级 JIT 模块 API
• jit/jit_test.go - 完整的测试套件和基准测试
• internal/torch/jit.go - 底层 CGO 绑定
• internal/torch/jit.h - C 函数声明
• lib/jit.cpp - 使用 libtorch 的 C++ 实现

修改文件：

• go.mod - 更新为 Go 1.25
• internal/torch/api.h - 添加 jit_module 类型
• internal/torch/exception.hpp - 添加 auto_catch_jit_module 模板
• lib/CMakeLists.txt - 添加 jit.cpp 到构建

---

Changes

New files:

• jit/jit.go - High-level JIT module API
• jit/jit_test.go - Comprehensive test suite with benchmarks
• internal/torch/jit.go - Low-level CGO bindings
• internal/torch/jit.h - C function declarations
• lib/jit.cpp - C++ implementation using libtorch
• JIT_IMPLEMENTATION_PLAN.md - Implementation documentation

Modified files:

• go.mod - Updated to Go 1.25
• internal/torch/api.h - Added jit_module type
• internal/torch/exception.hpp - Added auto_catch_jit_module template
• lib/CMakeLists.txt - Added jit.cpp to build

API Example

model, err := jit.Load("model.pt")
if err != nil {
    return err
}
defer model.Close()

model.Eval()

// Single output
output, err := model.Forward(input)

// Multi-output (e.g., BirdNET returns embeddings + predictions)
outputs, err := model.ForwardMulti(input, 2)
for idx, out := range jit.Outputs(outputs) {
    fmt.Printf("Output %d shape: %v\n", idx, out.Shapes())
}

Test plan

• Unit tests pass with BirdNET v3.0 model
• Tests cover: Load, LoadToDevice, Forward, ForwardMulti, iterator, eval/train modes, Close idempotency
• Concurrent inference test using testing/synctest
• Batch inference test
• Benchmarks: ~143-145ms/inference on CPU

🤖 Generated with Claude Code