Describe the bug
When trying to extend a CAGRA index that has less than 128 nodes, a segfault occurs.

Steps/Code to reproduce bug
Run the test BuildExtendSearch in cpp/test/neighbors/ann_cagra_c.cu and set main_data_size to 127

or (using CAGRA with C++ directly):
Change n_rows to 127 in one of the test inputs in ann_cagra.cuh

Expected behavior
The index gets extended successfully

Environment details (please complete the following information):
Bare-metal, compiled from source. I've seen this bug on CUDA toolkit 12.6, 12.5 and 12.4.