That's because you're rotating spherical creatures. A creature that's a line segment would be 42% longer when moving diagonally.
There's also that if you have creature like this:
_X
OOO
_V
rotating into
X_O
_O_
O_V
a dwarf could walk between it's arms diagonally.
Boats have same issue, 10x1 wall is water-tight in a line, but lets water though when rotated by 45°.
@Reelya: It uses ballastic formulas, as shown by tracing flying minecarts.
If it used a linear formula, then minecart sent even slightly airborne would never come back down until it hits something, which would have some novel applications but overall simulate things worse.
Also, it doesn't use a random roll. The moment a minecart tries to move into a tile where a dwarf stands there's a collision check, and dwarf has chance to dodge based on whether they're a dwarf and their dodging skill, and if collision succeeds the cart loses consistent amount of velocity, depending on what it hits.
I agree with @Egan_BW that at the very least other things could also have subtile positions kept to the hundreth thousand, like minecarts do, but I don't strongly see the need. It certainly wouldn't be very visible in gameplay, and would require rewriting and testing current movement code. I don't think this would have much of an impact in fps; creatures already count down when they've crossed over to next tile when walking.
I suspect that they meant something lot less complicated than wind (which isn't modeled beyond 10k speed reduction at too high speed in medium), or even subatomic scale simulation, though.
Making landforms into vectors internally, similarly to vector fonts, could also provide fps gains, I imagine, as one would have to check only the entire landform not its each individual tile.