improve compile time of evm.modifier.compose

[pfackeldey]

This PR replaces most python for-loops by jax.lax.scan loop constructs in evm.modifier.compose. It automatically transforms an array of evm.Modifier(s) into a single evm.Modifier of arrays. This rearrangement allows to use jax.lax.scan over the stacked axis.
Currently it is not clear how to do this for every type of evm.modifier.ModifierBase (e.g. evm.modifier.where, ..., and custom modifiers), however most modifiers in a typical analysis will be of type evm.Modifier and thus will use jax.lax.scan.
The compile time reduction is roughly a factor of 4-5, without any performance losses:

compile time: original
took 87.3687 s
eval time: original
took 0.1036 s
compile time: this PR
took 21.7520 s
eval time: this PR
took 0.1016 s

This benchmark has been performed with 103.000 nuisance parameters, of which:

• 100.000 are constrained by Poisson
• 1000 by log normals
• 1000 by Gaussians
• 1000 by shape (up/down templates) uncertainties

These number correspond to roughly 10x the amount of modifiers that are used in a typical HEP analysis.