Is your feature request related to a problem?
Like the Lucene Scalar Quantizer, add support for Faiss engine which accepts fp32 vectors as input without any range limitations and dynamically scalar quantize (recompute the min-max or quantiles and requantize data if required while merging segments or if the range of new data that will be ingested lies outside of existing min-max or quantile range) them into int8 vectors using non-uniform/training mechanisms without the need for any external quantization or extra training steps.

Also, for the disk based vector search use Faiss as the default engine (by replacing lucene engine) for x4 compression to keep it consistent with other compression levels.