Not just any: consider the dumbest possible cellular automate with 1 cell that always switches between black and white. Trivially reversible but mass is not conserved. It's easy to see that mass must oscillate though (and cannot grow infinitely), if the system is finite and time-reversible (and if not finite, then mass of the system is ill-defined).
Can we say that a CA is reversible if, in order to deduce the previous state, you must know the initial conditions? With your CA, if you did not know that the current state ultimately descended from one cell, you would not know whether the previous state involved two towers separated by an empty cell.
EDIT: I'm assuming that "neighbor" means a cell sharing an edge, but I realize that the Game of Life includes cells sharing a corner, so this probably isn't what you meant. Leaving it for... curiosity, I guess?
> In the reverse direction a horizontal row appears.
If you have a non-trivial row of live cells, then the cells immediately above that row will also become alive in the next (forward) instant. You end up with a row of N propagating upwards with a trail of rows of N-1.
I think your 1-cell example is actually a Garden of Eden in this rule -- there is no state that would produce it. The 1-cell itself isn't a still life, since as you note, it generates a vertical tower. Going backwards, a cell could only exist in the N+1'th state if it has a left neighbor in the N'th state; the only possible candidate for a predecessor would be the 2-cell row. But if we step this forward, we end up with three cells:
