Planned additions and improvements to NeuroTabModels.jl

[AdityaPandeyCN]

Multi AD Backend Support

• Add backend kwarg to NeuroTabRegressor / NeuroTabClassifier (enzyme, zygote, mooncake)
• Extend device kwarg to accept cpu, gpu, reactant
• Wire up Zygote extension
• Enzyme + Reactant path (device=reactant)
• Document device/backend compatibility

TabReD Temporal Embeddings

• Implement temporal embeddings per Temporal embedding
• Adapt constructor to support optional temporal embeddings
• Ensure compatibility with existing embedding options (piecewise, linear, etc.)

ModernNCA Core Model

• Implement ModernNCA per paper
• Benchmark against MLBenchmarks suite
• Code + package documentation

MLBenchmarks Datasets (TabReD)

• Add ingestion recipes for the 3 TabReD datasets under src/Datasets
• Add training scripts under benchmarks/
• Benchmark against Python TabReD reference implementation

[jeremiedb]

We can remove Mooncake for now. Some basic MN operators still seems unsupported (ex BatchNorm).
Enzyme on Cpu/GPU as well as Reactant, plus Zygote will be sufficient for now.