Make injection work more reliably on Windows

[ndmitchell]

Before 64/64 injection worked well, but nothing else really did. There are two root causes:

• The computed dll name when doing 64 -> 32 was computed as fsatrace6432.dll which was no good and didn't inject the hooks properly. I have fixed that up by rewriting the name mapping code.
• The spawning of fsatracehelper was failing because CreateProcess was calling back into inject, so it spawned 100 or so fsatracehelpers and then failed with the injection error from #28. I solve that with a static variable, so once we are actively trying to find out the fsatracehelper output we don't stop injecting. I guess this is a race condition? If so, a mutex could be used to prevent it? Or maybe its not? Certainly the new strategy works in practice.

With that, 64/64, 64/32 and 32/32 all work nicely. The 32/64 doesn't work, but reading https://stackoverflow.com/questions/494284/createremotethread-32-64-and-or-64-32 it doesn't seem it can without some seriously grim assembly. Given its a far less common operation, I'm happy to leave it out for now.

[ndmitchell]

With this final patch I can pass the full test suite with all the fixme's removed on Windows.

[ndmitchell]

I believe the Mac failure is spurious - I'd recommend clicking refresh in the CI.

[jacereda]

I haven't forgotten about this one, I want to fix some issues on the posix side before merging it.

[ndmitchell]

Thanks for the update, that's fine by me.

[jacereda]

With this final patch I can pass the full test suite with all the fixme's removed on Windows.

There're still a couple of them in the code (Arbitrary FSATest and allTests)

[jacereda]

I got a windows box, I'll try to figure out why the injecting test works.

[jacereda]

I'm merging this but I'm skeptical about the static variable. I've removed the injecting test and it still works for me.

[jacereda]

Thanks!