Bug in revolve implementation

[navjotk]

After a lot of testing, I continue to believe that our revolve implementation might not be doing the "right" thing. For 4 timesteps and 4 checkpoints, here is the schedule pyrevolve came up with:

Save
Forward from 0 to 1
Save
Forward from 1 to 2
Save
Forward from 2 to 3
Forward from 3 to 4
Reverse from 4 to 3
Load
Reverse from 3 to 2
Load
Reverse from 2 to 1
Load
Reverse from 1 to 0

Notice that there is no save between F(2, 3) and F(3, 4) and this is, I think, wrong. Even though we had 4 checkpoints, we wasted one by not storing this timestep.
Guillaume's utility, on the other hand, gave the following schedule for the same inputs (4 timesteps, 4 checkpoints):
Sequence: [WM_0, F_0->0, WM_1, F_1->1, WM_2, F_2->2, WM_3, F_3, B_4, RM_3, B_3, DM_3, RM_2, B_2, DM_2, RM_1, B_1, DM_1, RM_0, B_0, DM_0]

[navjotk]

I later discovered that this is because Revolve assumes the initial input is to be stored in the first checkpoint. For various reasons this is not necessarily true. We should be handling this a bit more smartly.