Optimise peak memory usage [4/4]

Summary:
**Why?**
Running analysis for multiple targets quickly increases the memory usage by buck2 daemon. Because of that glean indexing [bxl script](https://www.internalfb.com/code/fbsource//fbcode/glean/facebook/lang/clang/index.bxl) ooms sandcastle machines in some cases. The common issue is very high peak memory usage with a lot of allocations which are [not released immediately](https://fb.workplace.com/groups/starlark/permalink/1312921066063385/).

This change is to not create a new array in dfs traversal which creates a new copy every time using `nodes_to_visit[::stride]`. Instead, traverse same array in reverse order

Reviewed By: podtserkovskiy

Differential Revision: D68555015

fbshipit-source-id: f3c1ceb257638e21c0d6ba4b65b7491b665c635f

prelude/utils/graph_utils.bzl

@@ -237,7 +237,9 @@ def depth_first_traversal_by(
             fail("Expected node {} in graph nodes".format(node_formatter(node)))
         nodes_to_visit = get_nodes_to_traverse_func(node)
         if nodes_to_visit:
-            for node in nodes_to_visit[::stride]:
+            range_traversal = range(len(nodes_to_visit) - 1, -1, -1) if stride == -1 else range(len(nodes_to_visit))
+            for i in range_traversal:
+                node = nodes_to_visit[i]
                 if node not in visited:
                     visited[node] = None
                     stack.append(node)