Use inverse gamma distribution prior for sigma_alpha and sigma_beta

[erikbern]

• ~Also don't fit them in the log domain~ reverted this
• Also lower burnin again since the model seems a bit more robust now
• Also noticed CG is much faster than SLSQP

[erikbern]

Merging this for now since it works so much better, even though I'm not sure why