v0.20.0-pre.4

What's Changed

• Make TransformerEncoderLayer fields public by @Mnwa in #4053
• Feature muon by @NewBornRustacean in #3925
• Implement FromStr for Slice with parsing and error handling by @crutcher in #3983
• chore: Update to cubecl scalar refactor by @wingertge in #4062
• refactor: cubecl Runtime trait by @wingertge in #4065
• Fix scatter values backward by @khoek in #4064
• Refactor/autotuner by @nathanielsimard in #4068
• Fix MPS "Placeholder storage has not been allocated" error for embedding operations by @antimora in #4073
• Remove burn-import abstraction layer and use onnx-ir types directly by @antimora in #4033
• More correctness fixes in autodiff ops by @khoek in #4069
• Fix transaction read by @laggui in #4074
• Feat/error handling cubecl by @nathanielsimard in #4076
• Move types from burn-tensor by @laggui in #4050
• burn-store enhancements for troubleshooting and new enum skip flag by @antimora in #4051
• Re-enabled no-std support for safetensors store by @antimora in #4071
• Fix tch bf16 kind by @laggui in #4088
• Feat/runtime error by @nathanielsimard in #4079
• Fix ConstantOfShape output size determination by @antimora in #4085
• Fix reduce codegen to use turbofish for squeeze_dims by @antimora in #4086
• Fix Expand operation to use ONNX max-semantics by @antimora in #4082
• Add ONNX GridSample op support and tests by @antimora in #4084
• Fix Slice operation to handle empty ranges by @antimora in #4083
• Add RF-DETR model check for burn-import by @antimora in #4087
• Fix cubecl by @BjornTheProgrammer in #4092