perf: PQ performance

[BubbleCal]

Now we are optimizing the PQ performance, here to track all potential bottlenecks:

• [ ] 256bit/512bit SIMD for 4bit PQ
• [x] transforming: now we handle 4bit PQ case while transforming, which requires to collect the intermediate results
• [x] cache locality: the distance calculating can be optimized by better access pattern to distance table
• [x] constructing distance table: now it's 4x slower than computing distances
• [ ] find partitions: significant when nprobes is small

[wjones127]

Have we fixed either of these yet?

• [x] https://github.com/lancedb/lance/issues/2838
• [x] https://github.com/lancedb/lance/issues/2837

[BubbleCal]

Have we fixed either of these yet?

• [x] perf: presize KeepFiniteVectors transform #2838
• [x] perf: PQ transform re-allocates a lot #2837

not yet