Better to connect through diagonal corners. The stub shown is almost so short it's hard to fab, and the effective notch-down between trace and pad is particularly discontinuous.
But you'd only notice such discontinuity at, like, 40GHz.
It might not be pretty at all frequencies, but you have to look very close indeed (picosecond time scales) to tell for sure!
And at lower frequencies, it really doesn't matter, it just manifests as some lumped capacitance corresponding to pad area. Which is not much, and I guess you need extra capacitance anyway (for filtering?), of which the pad capacitance is merely a rounding error to the capacitor tolerance.
So, the situation is more or less irrelevant. Do what looks best!
Ethernet isn't very demanding on bandwidth: low 100s MHz, so stub lengths of some cm are acceptable; indeed, necessary -- consider the lengths of the PHY side windings of a typical 10/100 transformer. (Typical 1G transformers/PHYs however use a balanced driver, making the stub length irrelevant, nice.) Best practices are always nice of course, for when it does matter -- USB high speed modes, PCIe, etc.
Note that, most board-level high-bandwidth protocols, have mitigations in place: fairly generous thresholds (true of most digital logic, really), preshoot/overdrive, automatic frequency response compensation and echo cancellation, input stages with hysteresis and edge blanking (i.e. to ignore ringing following an edge), clock recovery/synchronization, error correction, etc. etc. Usually not all of these are applied, I think, but a good set of them, like DDR3 memory, or 4 or something (it's been a while since I looked it up TBH) has overdrive and edge blanking, as well as differential input thresholds (signals are single-ended, but with respect to a reference plane that's biased mid-supply; basically differential like LVDS, but all the '-' nodes are gathered together, saving on pins), and some automatic delay compensation to deal with trace length variation.
Tim