Add parallel option for sc.asd()

[cliffckerr]

No description provided.

[RomeshA]

Some options might include

• At each step, draw n proposed changes, try all of them, and pick the best one to use
• As above, but also update the parameter selection probabilities for the next generation based on the (otherwise unused) steps

[kelvinburke]

• Can use Parallelize: Send func, args and kwargs once to each process using pool initializer #621 if it gets implemented. We can keep the pool with each process having the function (the model) already passed to it - reducing the time for spin-up - only needing to pass the new parameters.
• Optionally can also use parallelize with a callback to process and suggest new samples as we get results one-by-one, rather than waiting for all n results.
• Could also try moving in the sum of improving directions (increase both A and B at the same time if increasing A and B individually was improving). This could be useful if we want to swap budget from one program to another, but reducing or increasing one parameter at a time is worse. Not sure how that would change the probabilities - maybe just it wouldn't change them.

[kelvinburke]

As an alternative, I have tend to found that starting multiple asds in parallel from different starting points tends to give the best results. So, if there was a built in way to do that, and possibly each asd at each step could choose from its own probabilities or perhaps some global combined probabilities and step sizes? Again not sure how that would really work - just putting ideas out there.