Summary

Add optional LZ4 compression to the .apr model format to reduce storage size and improve transfer speeds.

Motivation

• .apr model files can be large (MBs to GBs)
• Float32/float16 tensors compress 2-10x with LZ4
• Faster uploads to Hugging Face Hub
• Reduced disk usage for model storage
• Synergy with trueno-zram for runtime decompression

Proposed Changes

1. Add Compression enum to AprWriter/AprReader
2. Support LZ4 compression (via trueno's kernel)
3. Backward compatible: uncompressed files still work
4. Magic bytes indicate compression type

API Design

// Writing compressed model
let writer = AprWriter::new()
    .with_compression(Compression::Lz4)
    .create("model.apr")?;

// Reading auto-detects compression
let reader = AprReader::open("model.apr")?;

Integration

• trueno-gpu: GPU batch compression for large models
• trueno-zram: CPU SIMD compression fallback
• batuta: Orchestration for model export/import

Acceptance Criteria

• Compression enum: None, Lz4, Zstd
• AprWriter::with_compression() method
• AprReader auto-detects compression from header
• Backward compatible with uncompressed .apr files
• Benchmarks showing compression ratio and speed
• Tests for roundtrip compression