Improve large sync resumability

[nickvanw]

Summary

• checkpoint historical copy progress using LastKnownPk cursors
• preserve the VGTID position on copy-resume cursors so Vitess can run catchup before copying pk > LastKnownPk
• switch the connector read and cursor-peek paths from psdbconnect.Sync to direct vtgateservice.Vitess/VStream
• decode raw VStream FIELD, ROW, VGTID, LASTPK, and COPY_COMPLETED events in the connector, and reject copy cursors that lack LastKnownPk field metadata
• avoid truncating when state already has historical copy progress
• fail fast when record/truncate/schema-build sends fail so state cannot advance past data Fivetran did not receive
• propagate Update stream context through schema/shard/sync work and cancellation-aware timeout backoff

Testing

• GOCACHE=/tmp/go-build go test ./...
• GOCACHE=/tmp/go-build go vet ./...
• DATABASE_* go test -tags=integration -run TestIntegrationInterruptedCopyResume -count=1 -timeout 20m -v ./cmd/internal/server/handlers
• Earlier on this branch: GOCACHE=/tmp/go-build go test -race ./...
• Earlier on this branch: DATABASE_* INTEGRATION_STRESS=1 go test -tags=integration -run TestIntegration -count=1 -timeout 20m -v ./cmd/internal/server/handlers