DACs have different requirements depending on the application. You'll find high precision solutions using expensive components to get good absolute precision, but in audio you don't need that. Linearity is the most important parameter.

Since you'll be using an ATmega2560 which comes in a 100-pin package you can probably spare 16 I/Os, and then I would choose for parallel I/O. You'll have less timing problems than when working with serial I/O like I2S.

edit
Will seems to favor I2S and he has a point on condition that you want to connect to other, (semi-)professional audio equipment. As I understand it your digital audio path is very local, restricted to the connection between your AVR and the DAC. And then I would stick with the parallel. For 44.1kHz sample rate choose a crystal that allows you to get an interrupt every 22.676\$\mu\$s. Use the interrupt to automatically latch the last sample into the DAC, and then you have lots of time to prepare your next sample. That's the only timing you're dependent on. With I2S on the other hand you'll have more things to do. You may use SPI to shift the data out, but in the middle of a word you'll have to toggle the word clock. Things you don't have to worry about with parallel I/O.

The simplest ADC is an R-2R resistor ladder. The linearity depends on the resistors' matching, and there lies the problem for a discrete solution. You would need rather expensive precision resistors. The integrated approach is much better. The AD5547 also uses an R-2R ladder network (see fig. 17 on page 12), so why is this better than discretes? Integrated resistors (and capacitors) may have some tolerance on their nominal value, but their mutual match is the best you can get. So the 80k\$\Omega\$ resistors may be 80.5k\$\Omega\$, but you're sure they're all that value, and that's more important than the actual value itself.
The AD5547 accepts a 5V power supply and is 16-bit parallel in.

The application schematic is from the datasheet, so look there if this isn't very readable.

The DAC is registered, which means you can use the latch to update several DACs simultaneously, but you can make it asynchronous and transparent too by making/WR low and LDAC high (see page 13). Then the output will be continuously updated. But this is only possible if all inputs are updated simultaneously! With an 8-bit microcontroller you will probably write to 2 I/O ports, and then you'll need the latch!
Like most DACs it's current output, so you need to convert it to voltage. In this application with the AD8628 they use a RRIO (Rail-to-Rail I/O) opamp whose specs match the required resolution, but you can use another opamp if you like.
The ADR03 is a precision voltage reference, low drift and such, which for an audio application is overkill. Just make sure that your voltage reference is ripple and noise free!

edit
David rightly comments that this is somewhat older technology, but I thought it is OK since you're using a microcontroller (in lieu of a DSP) and a less-than-perfect class-D amplifier anyway. I picked this for my answer also because it is easy to understand.
David also mentions sigma-delta DACs, which have indeed better performance, but their working principle is a bit more complex, and they're (almost?) without exception serially controlled, for which the AVR doesn't have the hardware.

A buffer can serve many purposes in a circuit. There are other uses besides what I mention below, but these are some fairly common ones you might encounter.

One common use is as a time delay in data transmission. It can take a bit of time for a chip to recognize the change in a signal's voltage level and react to it (such as a digital HI to LO transmission). The amount of this slight propagation delay varies from one device to the next. Sometimes it is a few nano seconds, sometimes it is much more. Also, it does take some time for the signal to propagate through the copper from one part of the circuit board to another. When there are one or multiple data signals moving from one part of the circuit to another, it may be necessary to delay one so that they arrive in a specific order or at exactly the same time.

Buffers can be used as momentary data storage. Similar to the time delay, there are situations when you need a small amount of time to store a data bit such as pipelining, but this example can be a bit difficult to explain, depending upon your knowledge of computer architecture.

Another example is signal isolation. Because a buffer is essentially a signal repeater, it can be used to isolate the signal from two parts of the circuit. This is the specific use in your example, as noted by Ignacio in his answer. In this particular case, the DAC (digital to analog converter) is sensitive to loads. A load can be anything from a resistor, speaker, IC input, or even just a piece of wire. This "sensitivity to load" means the signal may change (distort) if too much or too little of a load is seen at the output of the DAC. To prevent this, a buffer chip is used to repeat the signal from the DAC output. The output of the buffer can be connected to whatever you want (within its own limitations) without affecting or distorting the output from the DAC in any way.

Lastly, a buffer can be used a signal amplifier. This is very similar to the previous example, but the reasons for using it can be different. While the output of some chip might not be extremely sensitive to loads, it may only be able to supply a few milliamps of current. In this case, you may want to use a buffer with a larger output current rating to amplifier the original signal so it can be used to drive some larger load.