With so little difference in VT (nominal) I don't think it
matters a whole lot. Either one is big enough that you
can throttle back several decades of current before
Vgs=0, so mirrors ought to be reliable-ish in matching.
A question is, does your CML perform any better (like
max well formed frequency vs current per gate) with
a current-source or a resistive tail characteristic?
The normal MOS will be less driven into saturation
at your setpoint current (or, will have to be wider).
If you need to run at a very low headroom then
lvt picks you up some tens of mV before you roll
over into linear region on the sources (sinks). The
question of which is better, remains.