internal/tags: use slices.SortFunc instead of sort.Sort

[th0114nd]

sort.Sort(t) requires boxing the TagSet into a sort.Interface, which
allocates on every call since TagSet doesn't fit in a single interface
word. slices.SortFunc is generic and takes the comparison function
directly, avoiding that allocation entirely.

Bumps the module's minimum Go version to 1.21, when the stdlib slices
package was introduced.

Co-Authored-By: Claude Sonnet 5 noreply@anthropic.com

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[th0114nd]

This change is breaking in that it relies on importers to upgrade to 1.21. The alternative is to add a dependency on golang.org/x/exp/slices.

All OSS dependents on gostats are forks of envoyproxy/ratelimit, and envoyproxy itself is at go1.26: https://github.com/envoyproxy/ratelimit/

https://pkg.go.dev/github.com/lyft/gostats?tab=importedby

[th0114nd]

Benchmark: allocation impact

Ran SerializeTags before/after this change (master @ 5f2727b vs this branch), -race=false to avoid race-detector overhead skewing results:

tags	before (allocs, B/op, ns/op)	after (allocs, B/op, ns/op)	delta
0	0, 0B, 3.75ns	0, 0B, 3.76ns	no change (untouched fast path)
1	1, 24B, 63.60ns	1, 24B, 63.22ns	no change (untouched fast path)
2	1, 48B, 89.17ns	1, 48B, 89.12ns	no change (untouched fast path)
4	1, 64B, 138.6ns	1, 64B, 138.2ns	no change (untouched fast path)
8	3, 392B, 256.1ns	2, 368B, 212.5ns	-1 alloc, ~17% faster
16	3, 760B, 665.9ns	2, 736B, 616.6ns	-1 alloc, ~7% faster

The 1-4 tag cases are unaffected as expected — they use hand-unrolled fast paths this change doesn't touch. The >4 tag default branch drops from 3 allocations to 2: pairs := make(TagSet, 0, numValid) (unchanged), sort.Sort(pairs)'s interface-boxing of the TagSet into a sort.Interface (removed — this is the allocation slices.SortFunc avoids), and the final output buffer (unchanged).

Methodology: go test -race=false -bench=BenchmarkSerializeTags -benchmem, GOFLAGS= to override the ambient -race default in this environment; each row averages b.N iterations via the standard testing.B loop.