Fix crash when loading old checkpoints with -reset_iterations 0 flag

[ChrisCummins]

Using -reset_iterations 0 when initialising training from a checkpoint made before commit 6990602 causes an error, as reported in issue #82. This patch adds a backwards compatibility fix.

For checkpoint files which do not contain the iteration counter, a warning is printed on checkpoint load and training continues as if not flag was given:

$ th train.lua $OPTIONS -init_from cv/checkpoint_10.t7 -reset_iterations 0
Running in CPU mode
Initializing from   cv/checkpoint_10.t7
reset_iterations: cv/checkpoint_10.t7 contains no iteration counter
Epoch 1.00 / 50, i = 1 / 17800, loss = 3.474951
Epoch 1.01 / 50, i = 2 / 17800, loss = 3.344084
...

The behaviour for checkpoints which do contain an iteration counter is unaffected.