storage: test that checkWriteRequest does not format metric families

[KR-Ravindra]

Problem

Since Pushgateway 1.11 switched to promslog, push latency grows with the size of the store as soon as any stored metric family has an inconsistent help string, even with --log.level=warn (#863: ~63% of CPU in slog.(*TextHandler).Handle under checkWriteRequest).

checkWriteRequest runs the consistency check on a scratch DiskMetricStore with promslog.NewNopLogger(). Up to prometheus/common v0.70.x that logger is a TextHandler writing to io.Discard whose Enabled() returns true, so the Info calls in processWriteRequest and GetMetricFamilies that pass whole *dto.MetricFamily values as attributes serialize every metric of the affected families on every push.

Fix

prometheus/common v0.71.0 (prometheus/common#960) makes NewNopLogger return a disabled slog.DiscardHandler. master picked that version up in #871, so this PR is now only the regression test: TestCheckWriteRequestInconsistentHelpAllocs pushes one metric against a stored 1000-metric family and asserts that an inconsistent help string does not cost more allocations than a consistent one. It fails on v0.69.0 (34115 extra allocations) and passes on master.

Fixes #863

This change was prepared with an AI agent operated by KR-Ravindra, who reviewed and tested it.

[KR-Ravindra]

Two weeks without a reviewer, so a short ping. This bumps prometheus/common to v0.71.0, which includes prometheus/common#960 where promslog.NewNopLogger returns slog.New(slog.DiscardHandler) instead of a handler that formats records and throws them away — pushgateway builds a nop logger per discarded push, so the formatting cost was real and measurable. Build and tests are green. Anything I can do to move it along?

[jan--f]

Thank you!