[STABLE ABI] Porting iir_cuda.cu

[pearu]

This issue collects tasks that block porting iir_cuda.cu to use torch stable ABI.

• expose AT_DISPATCH_FLOATING_TYPES to stable ABI, currently one need to implement the dispatch logic using switch block. Not a blocker but would nice to have.
• expose PackedTensorAccessor to stable ABI or replace its usage in iir_cuda.cu

^ @NicolasHug @scotts @janeyx99

[pearu]

All tasks here have resolutions in #4078.