perf: Optimize discovery.relabel by using label builder

[simonswine]

This avoids building labels unnecessarily and uses a retained LabelBuilder for relabelling itself.

Both cpu and allocations are reduced significantly:

benchstat before.txt after-v2.txt
name        old time/op    new time/op    delta
Relabel-12    10.2ms ± 1%     9.6ms ± 1%   -5.29%  (p=0.000 n=9+9)

name        old alloc/op   new alloc/op   delta
Relabel-12    9.36MB ± 0%    5.15MB ± 0%  -44.91%  (p=0.000 n=10+10)

name        old allocs/op  new allocs/op  delta
Relabel-12      168k ± 0%      132k ± 0%  -21.66%  (p=0.000 n=10+9)

Also in a real world deployment this is fairly impactful, both garbage collection and relabel use less cpu:

https://dev.grafana-dev.net/a/grafana-pyroscope-app/profiles-explorer?searchText=&panelType=time-series&layout=grid&hideNoData=off&explorationType=diff-flame-graph&var-serviceName=pyroscope-scrape-dev-001%2Falloy&var-profileMetricId=process_cpu:cpu:nanoseconds:cpu:nanoseconds&var-dataSource=edqsc0bzmfvnkc&var-filters=&var-filtersBaseline=&var-filtersComparison=&var-groupBy=&maxNodes=16384&from=now-30m&to=now&diffFrom=2025-02-04T14:30:26.865Z&diffTo=2025-02-04T14:45:22.388Z&diffFrom-2=2025-02-04T16:30:17.910Z&diffTo-2=2025-02-04T16:44:55.522Z&from-3=2025-02-04T14:00:00.000Z&to-3=2025-02-04T17:00:00.000Z&from-2=2025-02-04T14:00:00.000Z&to-2=2025-02-04T17:00:00.000Z

In terms of allocation bytes they are down by about 46%:

[simonswine]

cc @marcsanmi your Benchmark lives on