[Task] Provide unwrap_it const pointer overloads for half-precision floating point types #12585

Test files such as catch2_test_device_reduce.cu and catch2_test_device_scan.cu use half_t and nvbfloat16_t types, which are test-support wrappers around corresponding CUDA __half and __nv_bfloat16 types.

unwrap_t utility bridges these types and their corresponding CUDA's counterparts.

The implementation only handles conversions of unqualified pointers. Attempting to convert const half_t* encounters compiler error.

This gets in a way of adding warranted const qualifiers for input iterators in CUB algorithms.