Add utilities for basic fitting of chinchilla parametric scaling laws

[tyler-romero]

Fitting of scaling laws follows the original chinchilla methodology + known helpful improvements closely.

Important points:

1. It seems many scaling law papers dont do a full grid search over initialization params, which can lead to suboptimal fits.
2. Since chinchilla scaling laws are intended to fit the best-achievable loss for a given N, D or C, using overestimate_penalty=10 can be helpful to compensate against undertuned ladder rungs. Empirically I have found that using this leads to better predictions of held out points when doing a rollout over increasing N.
3. Bootstrapping is useful to get a feel for the plausible variance in predictions based on minor differences in underlying data but I would not consider it robust yet.