perf(tsdb): add targeted label cardinality scan

[yuchen-db]

Summary

MemPostings.Stats scans every label name and value while holding RLock, even when a caller only needs the highest-cardinality values for one label such as __name__. Periodic callers can therefore block new-series insertion on large postings indexes.

This adds TopLabelValuesBySeriesCount on MemPostings and Head. The method snapshots only the requested label's values and postings lengths under RLock, then performs top-K selection after releasing the lock. Existing Stats and /api/v1/status/tsdb behavior remain unchanged; Thanos must switch its periodic top-metrics collector to the new API in a follow-up.

Benchmark

The fixture contains approximately 1,000 __name__ values and 200,000 unrelated label values.

                         │   all-labels   │            single-label            │
                         │     sec/op     │   sec/op     vs base               │
PostingsCardinalityStats   18905.62µ ± 4%   30.03µ ± 3%  -99.84% (p=0.002 n=6)

                         │   all-labels   │            single-label             │
                         │      B/op      │     B/op      vs base               │
PostingsCardinalityStats   6326.86Ki ± 0%   24.23Ki ± 0%  -99.62% (p=0.002 n=6)

                         │   all-labels    │            single-label            │
                         │    allocs/op    │ allocs/op   vs base                │
PostingsCardinalityStats   200804.000 ± 0%   2.000 ± 0%  -100.00% (p=0.002 n=6)

Command:

go test ./tsdb/index -run '^$' -bench '^BenchmarkPostingsCardinalityStats$' -benchmem -count=6

Tests

• go test ./tsdb/index ./tsdb
• go test -race ./tsdb/index ./tsdb -run '^(TestMemPostingsTopLabelValuesBySeriesCount|TestHeadTopLabelValuesBySeriesCount)$' -count=1
• make lint with the repository-selected Go 1.23 toolchain

[PERF] TSDB: Add an API for computing top label values by series count without scanning unrelated labels.