[BUG] OOB access in multi-threaded CAGRA search

========= Invalid __global__ read of size 8 bytes
=========     at void cuvs::neighbors::cagra::detail::single_cta_search::search_kernel<(unsigned int)64, (unsigned int)64, (unsigned int)1, cuvs::neighbors::cagra::detail::dataset_descriptor_base_t<float, unsigned int, float>, unsigned int, cuvs::neighbors::filtering::none_sample_filter>(unsigned long, T4::DISTANCE_T *, unsigned int, const T4*, const T4::DATA_T *, const T4::INDEX_T *, unsigned int, const T5 *, unsigned int, unsigned long, const T4::INDEX_T *, unsigned int, T4::INDEX_T *, unsigned int, unsigned int, unsigned int, unsigned int, unsigned int *, unsigned int, unsigned int, unsigned int, T6)+0x40
=========     by thread (0,0,0) in block (0,15,0)
=========     Access to 0x6290000020 is potentially made before memory is allocated
=========     and is inside the nearest allocation at 0x6290000000 of size 64 bytes
=========     Saved host backtrace up to driver entry point at kernel launch time
=========         Host Frame: cuLaunchKernel_ptsz [0x3534c4] in libcuda.so.1
=========         Host Frame:  [0x14e7c] in libcudart.so.13
=========         Host Frame: cudaLaunchKernel_ptsz [0x572b7] in libcudart.so.13