Shared interner: fix leak of interned strings from scrape cache

[rfratto]

Ensures that interning and releasing strings are done in sync with the cache being manipulated to prevent duplicate interned references from being created.

Adds a metric to track how many interned strings exist to detect leaks. Also adds tests and a benchmark to keep track of this more easily.

name              old time/op    new time/op    delta
StringInterner-8     4.23s ± 0%     2.39s ± 0%  -43.51%

name              old alloc/op   new alloc/op   delta
StringInterner-8     421MB ± 0%     418MB ± 0%   -0.73%

name              old allocs/op  new allocs/op  delta
StringInterner-8     4.41M ± 0%     4.41M ± 0%   -0.00%

[rfratto]

I don't think this is right. This is releasing labels for series even before they go stale.

[rfratto]

I wrote some tests (which I'll push soon). This isn't "wrong," since there's three refs added to the interned string:

1. The series cache ref for a new series
2. The staleness ref for a new series
3. The staleness ref for an existing series

This will release the staleness ref correctly, though it spends a lot of unnecessary CPU time. Instead of doing this, the staleness ref should only be added for a new series and it should be removed when that series goes stale.

[rfratto]

Result of my benchmarks, scraping a target with 1,000,000 series 50 times:

name \ time/op    none.txt    orig.txt    old.txt     one-ref.txt
StringInterner-8  2.13s ± 0%  4.23s ± 0%  6.03s ± 0%   2.39s ± 0%

name \ alloc/op   none.txt    orig.txt    old.txt     one-ref.txt
StringInterner-8  396MB ± 0%  421MB ± 0%  397MB ± 0%   418MB ± 0%

name \ allocs/op  none.txt    orig.txt    old.txt     one-ref.txt
StringInterner-8  4.18M ± 0%  4.41M ± 0%  4.19M ± 0%   4.41M ± 0%

where:

• none.txt: no interning
• orig.txt: behavior of current release
• old.txt: behavior of the original status of this PR
• one-ref.txt: fixed PR that only tracks one reference per series and only interns when the series is added.

The newest approach is the best tradeoff between cpu/memory usage. I'll push a set of commits now.

[mattdurham]

LGTM