hotfix: limit concurrency of zfs send & recv commands

ATM, the replication logic sends all dry-run requests in parallel,
which might overwhelm the ZFS pool on the sending side.
Since we use rpc/dataconn for dry sends, this also opens one TCP
connection per dry-run request.

Use a sempahore to limit the degree of concurrency where we know it is a
problem ATM.
As indicated by the comments, the cleaner solution would involve some
kind of 'resource exhaustion' error code.

refs #161
refs #164

util/semaphore/semaphore.go

@@ -0,0 +1,38 @@
+package semaphore
+
+import (
+	"context"
+
+	wsemaphore "golang.org/x/sync/semaphore"
+)
+
+type S struct {
+	ws *wsemaphore.Weighted
+}
+
+func New(max int64) *S {
+	return &S{wsemaphore.NewWeighted(max)}
+}
+
+type AcquireGuard struct {
+	s        *S
+	released bool
+}
+
+// The returned AcquireGuard is not goroutine-safe.
+func (s *S) Acquire(ctx context.Context) (*AcquireGuard, error) {
+	if err := s.ws.Acquire(ctx, 1); err != nil {
+		return nil, err
+	} else if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	return &AcquireGuard{s, false}, nil
+}
+
+func (g *AcquireGuard) Release() {
+	if g == nil || g.released {
+		return
+	}
+	g.released = true
+	g.s.ws.Release(1)
+}