This commit changes the logic used for logprob inference. Instead of eager bottom-up conversion to measurable variables in the IR rewrites, we only convert nodes whose outputs were marked as "needs_measuring". This is achieved with the new `PreserveRVMappings.request_measurable` method.

This strategy obviates the need to undo unnecessary conversions. It also obviates a subtle need for graph cloning via the `ignore_logprob` helper, which prevented intermediate measurable rewrites from being reversed when they were needed to derive the logprob of valued variables, but were not directly valued. This indirect role of `ignore_logprob` is now done more explicitly and efficiently via the `request_measurable` method.

All other uses of `ignore_logprob` (and `reconsider_logprob`) were removed from the codebase

The `get_measurable_outputs` dispatching was also abandoned in favor of only considering outputs associated with value variables.

A new MergeOptimizerRewrite was written to further target local rewrites to only those nodes whose variables have been marked as `needs_measuring`.

Join/MakeVector/IfElse can output multiple interdependent variables. These are potentially measurable because in the logp each output is given a distinct value variable. However, this isn't known during the IR rewrites.

To circumvent this issue, we run an inner IR rewrite after giving dummy value variables to each output

Co-authored-by: Joseph Hall <joseph.hall@bp.com>

The True and False branches were being mixed up