Improve GECO

[dalessioluca]

what is the range of allowed value fo the GECO hyper-parameters (0; + infinity) ?
change of hyper-parameters should be proportional to the distance to the target

[dalessioluca]

87 "factor_nll_range" : [0.0, 1.0, 40.0], this should be allowed to become negative to prevent empty solution
88 "factor_sparsity_range" : [0.0, 1.0, 40.0], this should be stricntly positive

[dalessioluca]

I realized that sparsity should be controlled by fg_pixel_fraction (which should be removed from regularization dictionary).

Moreover the balance between reconstruction and KL should be that
a = b
Think about the edge cases when KL ->0
Therefore a,b should not go to zero.

maybe:
a = max(1, NLL_av /( NLL_av + KL_av) )
similarly for b

[dalessioluca]

In PR #20:

1. sparsity is adjusted based on fg_fraction
2. balance between KL and NLL is adjusted based on the NLL target

KL_LOGIT is normalized w.r.t to running average. The only question is wether the other KL should also be normalized in the same way......

[dalessioluca]

Interesting observation:
NLL is intensive and of order 1.
Sparsity is also intensive and of order 1.
If I switch off a foreground pixel I am going to pay a price in NLL but gain in SPARSITY....
It seems that these procedure will reach an equilibrium in which:
alpha = beta