mmaprototype: use meansLoad in MMARebalanceAdvisor

This commit switches to using meansLoad and passes the stores slice directly to
computeMeansForStoreSet. Once the other PR is merged and this one is rebased, we
should fix-up this commit to allocatorimpl: consult with mma for replicate queue
rebalancing.

Author: wenyihu6

pkg/kv/kvserver/allocator/mmaprototype/allocator_state.go

@@ -1305,7 +1305,7 @@ type MMARebalanceAdvisor struct {
 	// existingStoreSLS is the load summary for the existing store.
 	existingStoreSLS storeLoadSummary
 	// means is the means for the candidate set.
-	means meansForStoreSet
+	means meansLoad
 }
 
 // NoopMMARebalanceAdvisor is a no-op MMARebalanceAdvisor that always returns
@@ -1325,14 +1325,12 @@ func NoopMMARebalanceAdvisor() MMARebalanceAdvisor {
 func (a *allocatorState) BuildMMARebalanceAdvisor(
 	existing roachpb.StoreID, cands []roachpb.StoreID,
 ) MMARebalanceAdvisor {
-	var means meansForStoreSet
 	// TODO(wenyihu6): for simplicity, we create a new scratchNodes every call.
 	// We should reuse the scratchNodes instead.
 	scratchNodes := map[roachpb.NodeID]*NodeLoad{}
-	storeIDs := makeStoreIDPostingList(cands)
-	storeIDs.insert(existing)
-	means.stores = storeIDs
-	computeMeansForStoreSet(a.cs, &means, scratchNodes)
+	scratchStores := map[roachpb.StoreID]struct{}{}
+	cands = append(cands, existing)
+	means := computeMeansForStoreSet(a.cs, cands, scratchNodes, scratchStores)
 	// TODO(wenyihu6): pass in the actual ctx here
 	existingSLS := a.cs.computeLoadSummary(context.Background(), existing,
 		&means.storeLoad, &means.nodeLoad)