Burn 0.22.0 delivers 6.2x faster recompilation and a pure Rust stack

The Burn deep learning framework released version 0.22.0, removing generics from user code and replacing third-party backends with CubeCL. CNN recompilation dropped from 28.4 to 4.6 seconds, while adaptive memory pools cut peak training memory by 49%.

• CNN recompilation sped up 6.2x, from 28.4 to 4.6 seconds
• Adaptive memory pools cut peak CNN training memory by 49%
• ndarray and LibTorch backends replaced by unified CubeCL
• Stack moved to pure Rust: MLIR replaced by Pliron, SQLite by Turso