make batch prediction able to not retrain using predicted guesses

[ardunn]

right now, batch prediction will remove data points from all_xz_unsearched to simulate not having that data point before retraining n times for a batch size of n. So if predicting on 100 points, will predict point 1, recompute acquisition function for point 2, etc. the problem is recomputing the acquisition function is expensive for each one, especially if each is using a bootstrap (e.g., RF). However, there's no need for the model to actually be retrained for every batch; So maybe the model can be trained once per batch and the code can be rewritten to compute the acquisition function alone for each new point in the batch

[ardunn]

alternatively, just store the predictions temporarily and store them until the batch runs out

[ardunn]

can be done by separating the prediction of rocketsled.acq.acquire into two functions: train and acquire or moving all the training code into OptTask