My seat-of-the-pants understanding for load capacitors (corrections invited) goes like this:
When a crystal is cut for a certain load capacitance, it is measured with that capacitance across it during final factory trimming. There is nothing magical about the value. It is simply a way of saying, that if you design your circuit to present that same capacitance, then your crystal will be within the stated (.005% or whatever) tolerance.
So, you add up all the capacitance in your circuit, and then add in what's needed to bring it up to the spec. We'll use your numbers. The stray capacitance due to the traces on the board obviously will vary with the board, so let's guess 1.3 pf. A number I made up, to go with the capacitance of the microprocessor's oscillator, stated to be 1.7 pf. So, we've got 3 pf in parallel with the crystal. The crystal wants 18pf, so we have to make up the 15 pf difference with discrete parts.
Since the two load capacitors are in series (Gnd->cap->xtal->cap->Gnd), we double the cap value to 30pf. Two 30 pf caps in series give us the 15 pf we're looking for.
Note 1. I tried searching for typical PCB stray capacitance. It was all over the map. Suffice it to say, that as the hardware gets smaller, the capacitance will keep getting smaller. A lot of typical values claimed less than 1 pf.
Note 2. If there is more capacitance than spec, the crystal will oscillate at a lower frequency than specified. If there's less, then it's higher. You can see, that if you want to trim the oscillator to spec, it's easier to shoot for a lower capacitance and add some later, than to try the opposite.
Note 3. For fun, look up "gimmick capacitor".
Note 4. My "seat of the pants" explanation is sufficient as an introduction, and this technique works in many cases, but not everywhere. For a more in-depth look at the EE principles behind those capacitors, see this answer.
Crystal pins are not I/O pins, and 10pF is a guaranteed maximum not a nominal capacitance.
Usually 5pF will suffice for a crystal pin plus stray capacitance with reasonable layout (as short traces as reasonable, possibly over a ground plane). However, the AT90USB1286 datasheet does not recommend load capacitors of less than 12pF so 10pF total would seem to be marginal (you'd get 11pF if the input+stray was 5pF).
If you don't care about the exact frequency, 12pF or even a bit higher would probably be okay with a "10pF crystal", it would just shift the nominal oscillation frequency very slightly. Otherwise you can pick a different crystal.
Best Answer
The crystal in combination with the circuit on the microcontroller forms a crystal oscillator. The function of this circuit is to provide a clock for the microcontroller.
You could also use Spehro's Suggestion and use an external crystal oscillator. That combines a crystal and a circuit containing everything that's needed to make that clock signal.
It might be slightly cheaper to use a crystal instead of the crystal oscillator. However, you should follow the recommendations of the microcontroller's datasheet regarding that crystal, there mainly frequency is important. You should also follow the recommendations of the crystal manufacturer's datasheet, there the load capacitance is important.
It is not difficult to get this "right" but get it wrong and it just won't work and that will be a pain.
Also a parameter like the 20ppm accuracy is often irrelevant as crystals are by themselves already very accurate. Also the microcontroller itself doesn't care about accuracy, it would still work even if the clock is extremely inaccurate and varying over temperature and whatnot.