modify predict to allow fixed b, varying parameters

[bbolker]

based on this code

• we can probably allow fast=TRUE to handle the newparams argument
• we should fix b (by default?) by setting map when modifying parameters

[bbolker]

question for others (@mebrooks, @kaskr); does it make sense to always map b, or should we have a switch that enables this? (i.e., what are the reasonable use cases for someone wanting to (1) modify beta parameters and re-estimate b correspondingly vs. (2) modify beta parameters and leave b fixed at their MLE values? (option 2 seems more sensible to me but I don't know if I can rule out option 1)

[bbolker]

As suggested in recent issues around the marginaleffects package, we almost certainly do want to fix b when beta values are being modified (because e.g. things like log-likelihood profiling will break if we don't). The extend_re_form branch is relevant here too.