One idea I have is that i have implemented the input neurons incorrectly.
There is no single correct way to implement the solution to this problem. There are less efficient and more efficient ways for specific problems.
Use 18 input neurons and have the first 9 being the state before placing a move, and the next 9 being the state after placing a move.
This should work. The first set of 9 neurons will receive a representation of the current game state. The second set of 9 neurons will receive a representation of the intended action.
For predicting Q values from a state, action pair, this is a reasonable approach. Staying with the same architecture of 18 inputs, you could also have the second set of neurons be one-hot encoded to where the agent will to place the 'X' or 'O'.
In the game Tic Tac Toe however, you can look for more compact representations if you like. You have already noticed that because the game is simple and deterministic, that you can represent the agent's action as simply "desired next state". And in fact, the current state does not actually matter to a player, other than to enforce the rules of what are valid next states.
You won't be implementing the rules of the game into the agent - they are part of the environment. Therefore, you can do away with the initial state altogether in your estimate, and work with the end state of each move - this is called the afterstate in the RL literature, and you will find it used a fair bit in deterministic games, or even in non-deterministic games where the randomness happens before the action choice (e.g. in Backgammon). An afterstate representation is more efficient because it encodes the fact that you don't care what route a player took to get to certain board position, you just care about the value of that position as the game continues.
Having said all that, if your goal is to learn basic RL, then you don't need to be looking for the most efficient solution, just one that works. Don't expect your NN-based learner with state-action logic to be the most efficient learner however.
Another question I have in terms of when you teach the neural network, would you feed it the old state and then the state after both you and the opponent have made a move?
Not as inputs to the network at the same time, no. In your (state, action [=next_state]) representation, your action representation should be the board state for the current player's move and before the other player takes any action. The resulting next state however, will be after the other player takes their action.
If you want to train two separate bots against each other, then each would see the current state, then it would choose an action, then it would either get the reward for winning, or the opponent would take a turn. If the opponent won, then the first bot should receive the (negative) reward. If the opponent's move was not final, then the first bot should see the state after the opponent's move and get to choose its next action.
Again, this is inefficient. For a win/draw/lose game like Tic Tac Toe, you don't need two separate agents, each with their own learning algorithm, in order to train through self-play. You can instead alter Q-Learning slightly to work with the minimax algorithm. In brief this means alternating between the agent selecting actions that maximise the expected reward (for player 1) or minimise it (for player 2).
However, like before, your 2 networks set up should be able to work, and is quite interesting dynamic - you could try different learning parameters, different NNs etc, and see which learns to win quickly (but don't forget starting player has an advantage for early random play, so you'd want to switch which learning algorithm was used for which player to get a fair assessment).
The difference again is in terms of efficiency - a single network inside a modified RL with minimax will typically learn faster that two separate networks.
