From the datasheet:
"Analog Noise Canceling Techniques
Digital circuitry inside and outside the device generates EMI which might affect the accuracy of
analog measurements. If conversion accuracy is critical, the noise level can be reduced by
applying the following techniques:
a. Keep analog signal paths as short as possible. Make sure analog tracks run over
the analog ground plane, and keep them well away from high-speed switching digital
tracks.
b. The AVCC pin on the device should be connected to the digital VCC supply voltage
via an LC network as shown in Figure 24-10." [shows pin 44 with LC filter]
and
"Note: The same circuitry should be used for AVCC filtering on the ADC8-ADC13 side"
Guess I would recommend Arduino pins A0-A5 for analog readings.
Ok, you are saying that we could profit from a second filtering network on the other side of the IC.
But anyway: in the datasheet AVCC1 and AVCC are both called "AVCC". If they are internally connected as it looks, then L2 is shorted and the existing filtering network is screwed up. (I have no unmounted ATMEGA32U4 at hand to verify that.)
In the datasheet both are marked AVCC, both have the same potential. Because they are in opposite corners of the package I would rather apply the LC filter on both pins (44, 24).