Add a comment on readonly invertedlist and pin memory

namespace faiss {

/*

 * Use pin memory to build Readonly Inverted list will accelerate cuda memory copy, but it will downgrade cpu ivf search

 * performance. read only inverted list structure will also make ivf search performance not stable. ISSUE 500 mention

 * this problem. Best performance is the original inverted list with non pin memory.

 */

PageLockMemory::PageLockMemory(size_t size) : nbytes(size) {

    CUDA_VERIFY(cudaHostAlloc(&data, size, 0));

}