fix: Try a few slots when there's a buffer collision.

[kentquirk]

Which problem is this PR solving?

• Refinery's trace cache is implemented as a circular buffer. Until this PR, when the buffer wrapped around and tried to re-use a slot that was still in the cache, it would immediately eject it and replace it with the new trace. This PR changes that logic to try the next 5 slots, which means that the occasional long-lived trace won't be ejected just because it's next in line.

I benchmarked this and it's actually slightly faster than the old method, because it ejects traces less often.

For sites where only occasional traces live a long time, this should greatly decrease the number of times that a collect_cache_buffer_overrun occurs.

Short description of the changes

• Modify the Set() method to try several slots before giving up

[TylerHelmuth]

Are there any meaningful tests we can add to cache_test.go?