Fix buffer overflow in k-mer prefilter with highly conserved sequences

[mzueva]

CacheFriendlyOperations::findDuplicates underestimates output size when checking for buffer overflow. The check uses std::min(elementCount, currBinSize/2), assuming at most half of bin entries are duplicates. This assumption breaks when query k-mers are shared across a large fraction of the target database — as it happens with antibody variable region sequences, where conserved framework k-mers match ~70% of targets on consistent diagonals. The result is silent out-of-bounds writes to the foundDiagonals buffer, causing segfaults during prefiltering. With multiple threads, the corruption typically crashes at low progress (~8%).

To reproduce: run mmseqs easy-search with an antibody query against a multi-million antibody database (e.g. 10M clonotype sequences). Any dataset where conserved k-mers match a large fraction of targets will trigger it.

To fix: replace std::min(elementCount, currBinSize/2) with elementCount for a correct upper bound.