If null vectors are present than the non-CUDA path will include these in training

[westonpace]

I recently added null filtering to the IVF & PQ training in the CUDA path in https://github.com/lancedb/lance/actions/runs/10930801434/job/30344492517?pr=2910

I was curious what our non-CUDA path did. It does not fail but I believe it is using whatever garbage values are in the null vectors as real values and thus, we are training on more-or-less random data. If the null percentage is fairly low then this isn't really a big deal but it might be something worth fixing at some point.