Fix TCP relay source handover and shutdown races

[sylvesterkaczmarek]

Fixes #262. Related to #247 and #261.

Give each TCP relay instance its own cancellation context and join its listener/connection workers before Start returns, including after partial startup failure. Install sources in accept order before launching readers, close sockets when readers finish, and make source dialing context-aware.

Regression tests cover both source modes, failures in duplex/readable port setup, delayed reader scheduling, EOF cleanup, canceled dialing, and immediate port reuse after shutdown. The startup rollback, delayed reader, EOF, and canceled-dial regressions fail against unmodified main.

Validation

macOS arm64, Go 1.27.1:

• go test -race -count=1 -timeout=3m ./pkg/... passed independently and with Synchronize TCP relay tests with provider readiness #261.
• Lifecycle regressions passed 300 repetitions across -cpu=1,2,4, independently and with Synchronize TCP relay tests with provider readiness #261.
• go vet ./pkg/tcprelay and git diff --check passed.

Test limitation

A 90-run full relay stress test combined with #261 hit one TestRelayReadable startup failure (bind: address already in use). The existing test helper releases an ephemeral port before the provider binds it, leaving a separate port-allocation race. No race-detector warnings or lifecycle-regression failures were reported in that run. This runtime fix does not change test-port allocation.

This PR and #261 change disjoint files and target main independently.