Draft: Pipeline cuVS IVF-PQ artifact builds

[Xuanwo]

Summary

• Pipeline CPU-side artifact preparation with a bounded worker pool while keeping a single GPU scheduler.
• Avoid random training reads for cuVS IVF-PQ, fix strided tensor copies, and add optional cuVS RMM pool configuration.
• Add scheduling, GPU event, and drain timing metrics used to locate the remaining end-to-end bottlenecks.

Benchmark context

Dataset: paired-llama-3.2-1b-embeddings-lmsys-chat-1m, 100M sampled rows, IVF-PQ with num_partitions=256, num_sub_vectors=128, sample_rate=256, max_iters=50, and filter_nan=false. The GPU runs used a single device and did not run or count torch results.

On p4de.24xlarge local NVMe, the best validated cuVS end-to-end result was 358.32s; the same-machine full CPU baseline was 643.29s, about 1.80x faster for cuVS. Runs validated indexed_rows_verified=100000000, non-empty indices, and reasonable index bytes. An EBS comparison was much more IO-bound: cuVS 852.26s vs CPU 902.90s.

Draft status

This is a discussion/progress PR. It is not ready as a production-shaped final stack.

Known cleanup before merge:

• Replace benchmark-only env knobs such as LANCE_CUVS_PREPARE_WORKERS, LANCE_CUVS_SCAN_BATCH_READAHEAD, LANCE_CUVS_SCAN_FRAGMENT_READAHEAD, and LANCE_CUVS_RMM_POOL with supported configuration, stable defaults, or remove them.
• Remove benchmark instrumentation or convert it to tracing/metrics hooks.
• Re-run focused correctness and benchmark validation after cleanup.