Use `num_search_each_probe` for parallel evaluation of the objective

[taku-y]

Hi developers of combo,

Thanks for providing a great library. I want to use it for my projects.

I have a question about multiple evaluations at each probe. For a function f(x) which takes a long time to run, I want to compute f(x) for a number of x values in parallel. To test this, I ran tutorial.ipynb with the following modifications:

res = policy.random_search(max_num_probes=5, simulator=simulator(), 
                           num_search_each_probe=4)

# Originally, max_num_probes=80 and num_search_each_prob=1
# I expect that the total amount of computation is not changed
res = policy.bayes_search(max_num_probes=20, simulator=simulator(), score='TS', 
                          interval=5, num_rand_basis=5000, num_search_each_probe=4)

Is this the right way to do parallel computation as explained above?

And I'm concerned about actual computation time. The modified code seems slower
than the original, specifically on policy.bayes_search():

Original: 99.50355625152588 [sec]
Modified: 1856.6796779632568 [sec]

I expected the computation time is proportional to the number of evaluations of f(x),
so I don't understand the significant difference. That's why I'm wondering that I'm not correct
on using parallel evaluation.