Unstable target association when increasing `target_association_tolerance`

The current association rule of binding to the nearest disk center bears the potential instability of resulting in different association results for different values of target_association_tolerance.

For example, in the fig below, all evaluation points within the red interval have unstable QBX disk association. The left end of the interval is where the expanded disks meet on the boundary, and the right end is the point on the boundary that is equidistant from either expansion center.

When the target_association_tolerance is smaller than the depicted critical scenario, points from this interval are bound to the right disk whenever possible; as target_association_tolerance gets larger, the interval will start to partially turn to the left disk, and eventually be completely bound to the left disk.

One way to make the results stable could be to use ratios of the Euclidean distance over the original disk radius as the criteria for association, instead of plain Euclidean distance. The intuition behind it that, when considering a continuous expansion of disks with steadily increasing target_association_tolerance, every point gets bound to the disk that first touches it.

/cc @inducer @mattwala

Edit: re-upload the fig