Seems to me like it can.

I didn't actually read your solution, but it most likely suffers from usual hard case in Mo's algo which is erasing some value from the interval when you need to change borders.

The first solution that came into my mind is with maintaining some structure like deque <int> mp[MAX_A] and when you add some value X to the left you mp[X].push_front() and mp[x].push_back() if you add it to the right. To delete some value you pop_back() and pop_front() in a same way.

Also you store multiset <int> of values and after adding / deleting value you insert / erase

mp[X].back() - mp[X].front()

That's our bottleneck now, here's the log factor

To fix this log factor you should do it in a same way, you need to backtrack your changes instead of using set to keep them all.

There is a nice and short editorial with an example of code.