[Target association] "When in doubt, go with the QBX expansion" is a bad strategy when coupled with a non-zero target assoc tolerance

In !25 (comment 14062), @mattwala says:

Just to let you know: there's a difference in our target associattion code with the algorithm from the paper: it assigns centers to all targets for which it finds a close center, as opposed to only targets marked close to sources.

This is not a good approach when mixed with target association tolerance t > 0. Consider the following situation:

 |---------*-------|----> +x
source   center  targets

Going beyond the QBX expansion is a bad idea on either end, but going beyond the valid radius in the (above) +x direction leads to (IMO) avoidable problems. By that point, the FMM potential is bound to be quite good, so there's no reason to force the use of the (likely much worse) QBX potential.

This may be a contributing factor to #3 (closed).