feat(cudnn): add batch normalization forward/backward

[Rachit2323]

Summary

Batch normalization wasn't wrapped in the cudnn crate yet (it's marked as not done in features.md), so I added it. It's a pretty common thing to need for neural nets, so figured it'd be useful to have.

I followed the same pattern as the existing softmax module to keep things consistent.

Changes

• Added a new batchnorm module with these methods on CudnnContext:
  • derive_batch_norm_descriptor
  • batch_norm_forward_inference
  • batch_norm_forward_training
  • batch_norm_backward
• Added a BatchNormMode enum (per-activation, spatial, spatial persistent)
• Added a SupportedBatchNorm trait for f32 and f64
• Added a small TensorDescriptor::from_raw helper (pub(crate)) so the derived
  descriptor can be wrapped
• Marked batch norm as done in features.md

Testing

• cargo clippy -p cudnn passes
• cargo test -p cudnn passes
• Wrote a test that runs forward inference, forward training and backward on the GPU and checks the output is correct
• Ran it on Ubuntu 24.04, RTX 6000 Ada, CUDA 13 / cuDNN 9