Fix errant empty data frame after trailers

With the metadata batching changes recently, I introduced a bug whereby
the http2 stream encoder would end up not being able to set the close bit on a header frame, and instead placed it on a new data frame, which is illegal by the HTTP2 spec.

src/core/transport/chttp2/stream_encoder.c

@@ -122,6 +122,12 @@ static void begin_frame(framer_state *st, frame_type type) {
   st->output_length_at_start_of_frame = st->output->length;
 }
 
+static void begin_new_frame(framer_state *st, frame_type type) {
+  finish_frame(st, 1, 0);
+  st->last_was_header = 0;
+  begin_frame(st, type);
+}
+
 /* make sure that the current frame is of the type desired, and has sufficient
    space to add at least about_to_add bytes -- finishes the current frame if
    needed */
@@ -571,6 +577,7 @@ void grpc_chttp2_encode(grpc_stream_op *ops, size_t ops_count, int eof,
            a metadata element that needs to be unreffed back into the metadata
            slot. THIS MAY NOT BE THE SAME ELEMENT (if a decoder table slot got
            updated). After this loop, we'll do a batch unref of elements. */
+        begin_new_frame(&st, HEADER);
         need_unref |= op->data.metadata.garbage.head != NULL;
         grpc_metadata_batch_assert_ok(&op->data.metadata);
         for (l = op->data.metadata.list.head; l; l = l->next) {
@@ -580,9 +587,6 @@ void grpc_chttp2_encode(grpc_stream_op *ops, size_t ops_count, int eof,
         if (gpr_time_cmp(op->data.metadata.deadline, gpr_inf_future) != 0) {
           deadline_enc(compressor, op->data.metadata.deadline, &st);
         }
-        ensure_frame_type(&st, HEADER, 0);
-        finish_frame(&st, 1, 0);
-        st.last_was_header = 0; /* force a new header frame */
         curop++;
         break;
       case GRPC_OP_SLICE: