Lucene: using cagraHnswLayers > 1 and IVF-PQ with cagraIntermediateGraphDegree >= 128 results in extra slow performance

Reproduction steps:

1. run SearchScale benchmark script on the latest 26.08.0 main branches of cuvs and cuvs-lucene
2. Need to use a dataset with at least 5M vectors in order to trigger IVF-PQ
3. Need to use a dataset with a distribution that can handle setting cagraHnswLayers above 1 (e.g. 3)
4. run exactly two tests where for one of them you set cagraIntermediateGraphDegree=126 and another one where you set cagraIntermediateGraphDegree=128
5. Observe the results

Results:
cagraGraphDegree = 66, cagraIntermediateGraphDegree = 126:
Total Index Build Time: 580 seconds

cagraGraphDegree = 66, cagraIntermediateGraphDegree = 128:
Total Index Build Time: 1058 seconds

Note:
Reproducing this may require the PR to enable out-of-core index-building https://github.com/rapidsai/cuvs-lucene/pull/141
Alternatively, if your dataset can fit on your GPU, then this is reproducible on multiple GPUs (e.g. I have reproduced it on both an A10-G and an L40S)

Please see attachment for detailed metrics:

metrics_ySiC8I_GPU__10M-1536d__1-seg_3-hnsw-layers__corner-case-128.csv

Note: I do not see this to be reproducible when using NN_DESCENT or when setting cagraHnswLayers=1. This is an issue because cuVS limits us to cagraIntermediateGraphDegree <= 512, and sometimes we may need to set cagraHnswLayers > 1 to achieve sought recall. This is 100% reproducible

Start with the SearchScale benchmark script against the latest cuvs and cuvs-lucene branches, using a 5M+ vector dataset and cagraHnswLayers=3. Compare cagraIntermediateGraphDegree values 126 and 128, using PR 141 for out-of-core setup if needed; done means identifying and correcting the reproducible build-time regression while preserving the reported configurations.