Address review feedback on the fp32-cast: instead of silently casting any
non-bf16/fp16 dtype (which would also accept int/bool), reject non-floating
inputs fast with TypeError and only cast floating dtypes (fp32/fp64) to bf16.
Extract the logic into a module-level _coerce_fp4_input_dtype() shared by
quantize_input() and apply(), and add a CPU-only unit test (no flashinfer/CUDA)
covering bf16/fp16 passthrough, fp32/fp64 -> bf16, and int/bool rejection.
Raghav
91f2acbbc9