Remove host-fp32 → device-fp16 downcast path in NN-Descent once cuML accepts fp16 natively

[jinsolp]

Background

NN-Descent currently supports internal_distance_dtype = CUDA_R_16F for fp32 inputs.
When the input lives on host memory, we downcast it to fp16 while copying it to device, so the on-device copy is in fp16 instead of its native fp32 dtype.
Doing this in NN-Descent is a special case relative to the rest of cuVS.

Proposal

Once cuML UMAP and HDBSCAN can natively accept fp16 input (tracked in NVIDIA/cuml#8102), remove the host-fp32 → device-fp16 downcast path from NN-Descent:

• Drop the d_data_half_ buffer and convert_copy_kernel from cpp/src/neighbors/detail/nn_descent.cuh.
• Remove the internal_distance_dtype parameter from the C++, C, and Python index params.

Related

• Originally introduced in PR Preserve input memory location for NN Descent #1928