Decouple vector store and encoders, also implementing openai ada support

[maxyu1115]

Decouple vector store and encoders, also implementing openai ada support.

Will not pursue using ada yet, it is too much slower than USE due to network latency. Here's the results of my experiment, ran on 10 wikipedia pages.

Splitting 11 documents into 3121 sentences took 0.033806s
[ada] Single: total 504.94618s, avg 0.16178986863184877s per item
[use] Single: total 3.12516s, avg 0.0010013329061198334s per item
[ada] 5 batch: total 141.952905s, avg 0.2274886298076923s per item
[use] 5 batch: total 1.017821s, avg 0.0016311233974358975s per item
[ada] 10 batch: total 76.166642s, avg 0.24412385256410254s per item
[use] 10 batch: total 0.732307s, avg 0.0023471378205128205s per item
[ada] 20 batch: total 44.71862s, avg 0.2866578205128205s per item
[use] 20 batch: total 0.530946s, avg 0.0034035000000000003s per item
[ada] 50 batch: total 24.703333s, avg 0.39844085483870967s per item
[use] 50 batch: total 0.431493s, avg 0.006959564516129032s per item
[ada] 100 batch: total 19.059711s, avg 0.6148293870967741s per item
[use] 100 batch: total 0.397267s, avg 0.012815064516129031s per item