feat: Add torch.compile support for ~2x faster inference

[fspecii]

Summary

This PR adds support for PyTorch 2.0+ torch.compile to significantly improve inference speed (~2x on supported GPUs like RTX 4090, A100).

Changes

New Environment Variables

• HEARTMULA_COMPILE: Enable torch.compile (true/false, default: false)
• HEARTMULA_COMPILE_MODE: Compile optimization mode (default/reduce-overhead/max-autotune)

Code Changes

• Added apply_torch_compile() function that compiles the model backbone and decoder
• Auto-detects Triton availability and falls back to eager backend if not found
• Updated create_quantized_pipeline() with compile parameters
• Applied torch.compile to both quantized and non-quantized pipelines
• All configurations (single GPU, multi-GPU, sequential offload) now support torch.compile

Documentation

• Updated .env.example with new configuration options
• Added torch.compile section to README.md with usage examples

Usage

# Enable torch.compile
HEARTMULA_COMPILE=true python -m uvicorn backend.app.main:app --host 0.0.0.0 --port 8000

# With max performance mode
HEARTMULA_COMPILE=true HEARTMULA_COMPILE_MODE=max-autotune ./start.sh

Requirements

• PyTorch 2.0+
• Linux/WSL2: pip install triton
• Windows: pip install -U 'triton-windows>=3.2,<3.3'

Notes

• First generation will be slower due to compilation
• Subsequent generations benefit from compiled kernels
• Falls back gracefully if compilation fails

Related

This implements a feature similar to heartlib PR #64.