CKSEL typically stands for clock select. There are several choices that can be selected via an internal multiplexer-- some for test, others for alternate sources.
Since you are using a xtal, you can provide the code to select that option: which are in the first 2 rows of p 26.
You can use 3 bottom rows of table 4 for a crystal. Each is tuned to a particular range of frequencies as shown.If your crystal reference was design to oscillate around 4MHz, you would set the sel bits to code, 111, for example. 16Mhz only works with CKOPT low using any of the three select codes shown in row 4 of the table.
External caps should be 12-22pF not 22uF for the xtal according to the same table. It also only specifies a max oscillator limit of 16MHz, so I would not use a 20MHz xtal.
Looks to me like the watchdog timer functionality is internal.
*note xtal and crystal are used synonymously here.
Your placement is fine.
Your routing of the crystal signal traces is fine.
Your grounding is bad. Fortunately, doing it better actually makes your PCB design easier. There will be significant high frequency content in the microcontroller return currents and the currents thru the crystal caps. These should be contained locally and NOT allowed to flow accross the main ground plane. If you don't avoid that, you don't have a ground plane anymore but a center-fed patch antenna.
Tie all the ground immediately associated with the micro together on the top layer. This includes the micro's ground pins and the ground side of the crystal caps. Then connect this net to the main ground plane in only one place. This way the high frequency loop currents caused by the micro and the crystal stay on the local net. The only current flowing thru the connection to the main ground plane are the return currents seen by the rest of the circuit.
For extra credit, so something similar with the micro's power net, place the two single feed points near each other, then put a 10 µF or so ceramic cap right between the two immediately on the micro side of the feed points. The cap becomes a second level shunt for high frequency power to ground currents produced by the micro circuit, and the closeness of the feed points reduces the patch antenna drive level of whatever escapes your other defenses.
For more details, see https://electronics.stackexchange.com/a/15143/4512.
Added in response to your new layout:
This is definitely better in that the high frequency loop currents are kept of the main ground plane. That should reduce overall radiation from the board. Since all antennas work symmetrically as receivers and transmitters, that also reduces your susceptibility to external signals.
I don't see the need to make the ground trace from the crystal caps back to the micro so fat. There is little harm in it, but it is not necessary. The currents are quite small, so even just a 8 mil trace will be fine.
I really don't see the point to the deliberate antenna coming down from the crystal caps and wrapping around the crystal. Your signals are well below where that will start to resonate, but adding gratuitous antennas when no RF transmission or reception is intended is not a good idea. You apparently are trying to put a "guard ring" around the crystal, but gave no justification why. Unless you have very high nearby dV/dt and poorly made crystals, there is no reason they need to have guard rings.
Best Answer
Calibrating against the mains frequency, as Tony suggests, is a bad idea. Long-time accuracy may be good, short-time accuracy isn't.
edit
Tony is dismissive about my reference, but that's no problem, there are other sources which confirm this. (Note that he does use my reference to show an absolute accuracy of 10 mHz/50 Hz = 0.1 ppm (sic). It looks like he is so preoccupied with his 10\$^{-10}\$ that he doesn't see a factor thousand error.) Maybe he accepts the authority of the ENTSOE, that's the "European Network of Transmission System Operators for Electricity". They should know. From this document:
This site gives you a real-time view of the deviation.
Even if we ignore the 200 mHz incidents there are still the 20 mHz deviations. We're talking about 400 ppm, that's more than an order of magnitude than the error of the uncalibrated crystal. 4000 ppm or two orders of magnitude taking the reference incidents into account. So the conclusion remains the same: the line frequency's short-term accuracy is by no means good enough to calibrate a crystal.
end of edit
The graph shows that a 50Hz mains frequency continuously fluctuates between 49.9Hz and 50.1Hz, that's a 0.2% error, or 2000ppm. An uncalibrated watch crystal is 20ppm accurate. (Horizontal scale is days.)
This device may be of help:
It's a Chip Scale Atomic Clock which outputs a 10MHz square wave with 1.5 \$\times\$ 10\$^{-10}\$ accuracy, several orders of magnitude more accurate than TCXO (Temperature Controlled Crystal Oscillator). Tune your oscillator so that you get 10 000 000 pulses from the CSAC over 32 768 cycles of your crystal.
Only 1500 dollar, which sounds like a bargain to me. (Your own fault, you should have mentioned a budget :-))
edit
Cheaper? OK, this OCXO (Oven Controlled Crystal Oscillator) has 5ppb (0.005ppm) frequency stability and less than 0.1ppm aging per year. About 150 dollar. Available in 16.384MHz, which is a multiple of 32.768kHz (500x). You mentioned this in your question, though there's really no reason for this.
Some GPS receivers have a 1 PPS (Pulse Per Second) output, which should have high accuracy as well. You would have to count cycles of your own 32.768 kHz clock over at least 30 seconds to get at 1 ppm accuracy. Ideally a single second will get you 32 768 counts \$\pm\$1 count, which is only a 30 ppm resolution.