Low-hanging fruit in vector.Read (#4881)

* Read directly from segmap rather than using `ReadBytes`. 
* Include a count in each segment (this is a breaking change to vng). Use this to pre-size slices.
* Reuse a single buffer across all internal read calls.
* Avoid unnecessary copies for string/bytes vectors.
* Correctly handle dict vectors, now that they're enabled.
* Add a benchmark tracking vng read performance and comparing it to zng and raw varints.

On wrccdc_mono_10m.vng this cuts vector.Read time from 270ms -> 140ms for id.resp_p and 470ms -> 180ms for uid.

runtime/op/meta/vec_sequence.go

@@ -120,11 +120,7 @@ func newVecSequenceScanner(ctx context.Context, zctx *zed.Context, pool *lake.Po
 	if err != nil {
 		return nil, ksuid.KSUID{}, err
 	}
-	vngReader, err := vng.NewReader(object)
-	if err != nil {
-		return nil, ksuid.KSUID{}, err
-	}
-	vectorReader := vector.NewReader(vngReader, demandOut)
+	vectorReader := vector.NewReader(object, demandOut)
 	scanner, err := zbuf.NewScanner(ctx, vectorReader, nil)
 	if err != nil {
 		return nil, ksuid.KSUID{}, err

vector/materializer.go

@@ -217,6 +217,15 @@ func (v *constants) newMaterializer() materializer {
 	}
 }
 
+func (v *dict) newMaterializer() materializer {
+	var index int
+	return func(builder *zcode.Builder) {
+		tag := v.tags[index]
+		builder.Append(v.dict[tag].Value.Bytes())
+		index++
+	}
+}
+
 func (v *maps) newMaterializer() materializer {
 	var index int
 	keyMaterializer := v.keys.newMaterializer()
@@ -268,6 +277,20 @@ func (v *records) newMaterializer() materializer {
 	}
 }
 
+func (v *sets) newMaterializer() materializer {
+	var index int
+	elemMaterializer := v.elems.newMaterializer()
+	return func(builder *zcode.Builder) {
+		length := int(v.lengths[index])
+		builder.BeginContainer()
+		for i := 0; i < length; i++ {
+			elemMaterializer(builder)
+		}
+		builder.EndContainer()
+		index++
+	}
+}
+
 func (v *unions) newMaterializer() materializer {
 	var index int
 	payloadMaterializers := make([]materializer, len(v.payloads))