Reduce session expiry scans and add an optional admission cap

[linkdata]

Closes #396.

Changes

• Sweep expired Sessions every tenth maintenance pass. Request retirement still runs every pass, and Session.Close unregisters immediately.
• Add an opt-in global Jaws.MaxSessions cap. Admission refuses new Sessions at the limit without evicting existing ones; expired Sessions count until cleanup.
• Add an opt-in Jaws.MaxSessionsPerIP cap on all registered Sessions per client address bucket. SessionMiddleware returns HTTP 429 when that bucket is full and global capacity remains. Existing Sessions continue to work.
• Return HTTP 503 from SessionMiddleware when the global cap or another creation failure prevents a required Session. NewSession returns nil without closing an existing cookie Session when admission or cookie publication fails.
• Document that AutoSession may leave a Request without a Session at the cap; use the middleware where a Session is required.

Both caps default to zero (disabled). Set the per-IP cap below the global cap to reserve capacity for other address buckets. NewSession needs a free slot under each enabled cap to rotate a Session; a refused rotation retains the old Session. Visitors in one bucket share a quota, while clients using multiple addresses can still fill the global cap. Rate-limit Session-creating routes at the proxy.

Verification

• go generate ./... (no generated changes)
• go vet ./..., staticcheck ./..., golangci-lint run (via go run ...@v2.14.0), gosec -quiet ./..., go build ./..., gofumpt -l
• JAWS_REQUIRE_NODE=1 go test -race ./... and JAWS_REQUIRE_NODE=1 go test ./...
• Focused race tests for capacity refusal, HTTP 429/503, cookie publication, address buckets, rotation at the per-IP cap, and ten-tick expiry sweeping
• Root-package race coverage: deleteSessionIfCurrentLocked 100% (99.7% package-wide)

Benchmark

BenchmarkJawsMaintenanceSessions measures sequential maintenance passes with 10,000 live Sessions. Apple M5 Max, -count=6 -benchtime=300ms -cpu=1, compared with benchstat:

Before	After	Change
182.11 µs/op	17.69 µs/op	−90.28% (p=0.002)