Improving Dual-Encoder Training through Dynamic Indexes for Negative Mining