I don't think the autodock units are "broken" so they can't be "fixed". The Vina work uses a different computation method that just happens to be extremely efficient under Linux. Could Autodock be optimised better? Possibly. Though it's not as simple as just setting a couple of flags and recompiling. Windows and Linux use different memory structures so the code would need to be optimised to make suit while still producing consistent results.
As for speed, Linux is typically faster on every project, but in some cases it's only barely and the difference is less than the performance increase achieved by switching from a mechanical HDD to a SSD. HCC and FAHV are the stand out examples, the latter being SO much better it pushes Linux machines' points output up significant'y compared to Windows machines on the same project. Most project still award more points to Windows machines, however, due to the disparity between BOINC benchmarks on Windows and Linux. There has been significant discussion about that in the past but I don't think it's ever been resolved.