instead of finding and copying the dll each time a binary is produced,
the path of the produced binary is stored in a unique file (to avoid race
condition) and the target for that binary is added as a dependency to a new
target named `bundle` that does the finding and the copying once for all.

this avoids race condition while building targets in parallel, workaround
the inability of cmake to track a common list between all targets and to keep
it from configure to build step

this makes the bundle process more reliable and avoid things being copied
twice. Before that I've seen copy processes starting to copy the same file
even with the -n option telling to not copy if the file exists, then seeing
one of them failing because they both started when the file was not yet
copied but one finished the copy before the other one

this also makes the bundling faster by only doing things once for all

the dll are then moved to a lib/ subdirectory inst...