dev_refthread: Do not initialize *ref when reference was not acquired

[cem]

Summary

Like the companion API devvn_refthread, leave *ref uninitialized when a
reference was not acquired. Initializing to 1 provides a vaguely
correct-looking but bogus value for broken callers to (mistakenly) pass to
dev_relthread() when refthread fails.

Make it even more clear to consumers that dev_relthread is only valid when
dev_refthread succeeds.

Test Plan

I have some other changes I want to make in this area, but this is the only one
that changes API (slightly). I'll include the same reviewers on those future
patches as well.

[kib]

I do not object but I must note that unconditional *ref = 1 looks strange to me. And since I am the author of that line, I believe that I added it that way for some reason which I cannot remember now. Most likely it was a compatibility issue with some third-party code. At least I looked at all in-tree callers and they looks fine.

[cem]

In D16885#359928, @kib wrote:

I do not object but I must note that unconditional *ref = 1 looks strange to me. And since I am the author of that line, I believe that I added it that way for some reason which I cannot remember now. Most likely it was a compatibility issue with some third-party code. At least I looked at all in-tree callers and they looks fine.

Maybe the atomic_add_long was in a different place at the time. I did not investigate history thoroughly. Thanks for the quick review!