Port HEBO into PROTEIN

[ClashLuke]

You likely can't use this as-is, but this is the backported version of the PROTEIN I'm currently running locally. It significantly outperforms PROTEIN (stronger GP and sampling) and HEBO (cost aware).

Specifically, PROTEIN's failure mode is that it loves boundaries. The core issue is that PROTEIN always picks the argmax, while HEBO picks the argmax every third sample with the other two being exploratory. HEBO therefore does not run into this.

[sozforex]

Compared sweeps for breakout between 4.0 [69fcbcf] and this PR [on top of 69fcbcf] using a bit narrower sweep config than the default.

Run# of second runs reaching specified perf for each algo:

perf	protein run#	CL_protein run#
>0.5	81	79
>0.6	81	107
>0.7	118	129
>0.8	166	143
>0.9	188	165
>0.93	205	171
>0.95	224	180
>0.96	308	229
>0.97	308	251
>0.98	314	255
>0.985	353	299
>0.99	459	373
> 0.995	473	?

sweep on this branch stopped on 427th run with
RuntimeWarning: overflow encountered in exp
at line 843 in sweep.py:
gp_c = np.exp(gp_log_c)

[ClashLuke]

awesome, thank you for running this!