Add "tiers" to data_index (#1147)

This is the first step in moving our `compute` infrastructure from a
recursive model to a looped one, which is potentially faster to compile
and requires less memory.

For now all this does is automatically assign a "tier" to each
computeable quantity that indicates how "deep" in the dependency tree
that quantity lies.

tier of 0 means no dependencies on other data, but may depend on
transforms, profiles etc.
tier of 1 means it depends on only tier 0 stuff,
tier of 2 means it depends on tier 0 and tier 1, etc etc.

The goal is that if you were to compute things in a loop with the order
determined by the tiers, then it would always compute dependencies in
the correct order.

desc/compute/__init__.py

@@ -94,3 +94,30 @@ def _build_data_index():
 
 
 _build_data_index()
+
+
+def set_tier(name, p):
+    """Determine how deep in the dependency tree a given name is.
+
+    tier of 0 means no dependencies on other data,
+    tier of 1 means it depends on only tier 0 stuff,
+    tier of 2 means it depends on tier 0 and tier 1, etc etc.
+
+    Designed such that if you compute things in the order determined by tiers,
+    all dependencies will always be computed in the correct order.
+    """
+    if "tier" in data_index[p][name]:
+        return
+    if len(data_index[p][name]["full_with_axis_dependencies"]["data"]) == 0:
+        data_index[p][name]["tier"] = 0
+    else:
+        thistier = 0
+        for name1 in data_index[p][name]["full_with_axis_dependencies"]["data"]:
+            set_tier(name1, p)
+            thistier = max(thistier, data_index[p][name1]["tier"])
+        data_index[p][name]["tier"] = thistier + 1
+
+
+for par in data_index.keys():
+    for name in data_index[par]:
+        set_tier(name, par)