Add more avg_over_time test cases with extreme values

These tests were initially created by @crush-on-anechka. I modified
them slightly.

Signed-off-by: beorn7 <beorn@grafana.com>

promql/promqltest/testdata/functions.test

@@ -1013,6 +1013,79 @@ eval instant at 1m sum_over_time(metric[2m])
 eval instant at 1m avg_over_time(metric[2m])
   {} 0.5
 
+# More tests for extreme values.
+clear
+# All positive values with varying magnitudes.
+load 5s
+  metric1 1e10 1e-6 1e-6 1e-6 1e-6 1e-6
+  metric2 5.30921651659898 0.961118537914768 1.62091361305318 0.865089463758091 0.323055185914577 0.951811357687154
+  metric3 1.78264e50 0.76342771 1.9592258 7.69805412 458.90154
+  metric4 0.76342771 1.9592258 7.69805412 1.78264e50 458.90154
+  metric5 1.78264E+50 0.76342771 1.9592258 2.258E+220 7.69805412 458.90154
+
+eval instant at 55s avg_over_time(metric1[1m])
+  {} 1.6666666666666675e+09
+
+eval instant at 55s avg_over_time(metric2[1m])
+  {} 1.67186744582113
+  
+eval instant at 55s avg_over_time(metric3[1m])
+  {} 3.56528E+49
+  
+eval instant at 55s avg_over_time(metric4[1m])
+  {} 3.56528E+49
+  
+eval instant at 55s avg_over_time(metric5[1m])
+  {} 3.76333333333333E+219
+  
+# Contains negative values; result is dominated by a very large value.
+load 5s
+  metric6 -1.78264E+50 0.76342771 1.9592258 2.258E+220 7.69805412 -458.90154
+  metric7 -1.78264E+50 0.76342771 1.9592258 -2.258E+220 7.69805412 -458.90154
+  metric8 -1.78264E+215 0.76342771 1.9592258 2.258E+220 7.69805412 -458.90154
+  metric9 -1.78264E+215 0.76342771 1.9592258 2.258E+220 7.69805412 1.78264E+215 -458.90154
+  metric10 -1.78264E+219 0.76342771 1.9592258 2.3757689E+217 -2.3757689E+217 2.258E+220 7.69805412 1.78264E+219 -458.90154
+
+eval instant at 55s avg_over_time(metric6[1m])
+  {} 3.76333333333333E+219
+  
+eval instant at 55s avg_over_time(metric7[1m])
+  {} -3.76333333333333E+219
+  
+eval instant at 55s avg_over_time(metric8[1m])
+  {} 3.76330362266667E+219
+
+eval instant at 55s avg_over_time(metric9[1m])
+  {} 3.225714285714286e+219
+
+# Interestingly, before PR #16569, this test failed with a result of
+# 3.225714285714286e+219, so the incremental calculation combined with
+# Kahan summation (in the improved way as introduced by PR #16569)
+# seems to be more accurate than the simple mean calculation with
+# Kahan summation.
+eval instant at 55s avg_over_time(metric10[1m])
+  {} 2.5088888888888888e+219
+
+# Large values of different magnitude, combined with small values. The
+# large values, however, all cancel each other exactly, so that the actual
+# average here is determined by only the small values. Therefore, the correct
+# outcome is -44.848083237000004.
+load 5s
+  metric11 -2.258E+220 -1.78264E+219 0.76342771 1.9592258 2.3757689E+217 -2.3757689E+217 2.258E+220 7.69805412 1.78264E+219 -458.90154
+
+# Thus, the result here is very far off! With the different orders of
+# magnitudes involved, even Kahan summation cannot cope.
+# Interestingly, with the simple mean calculation (still including
+# Kahan summation) prior to PR #16569, the result is 0. That's
+# arguably more accurate, but it still shows that Kahan summation
+# doesn't work well in this situation. To solve this properly, we
+# needed to do something like sorting the values (which is hard given
+# how the PromQL engine works). The question is how practically
+# relevant this scenario is.
+eval instant at 55s avg_over_time(metric11[1m])
+  {} -1.881783551706252e+203
+# {} -44.848083237000004 <- This is the correct value.
+
 # Test per-series aggregation on dense samples.
 clear
 load 1ms