MTU: maximum size of IP packet (including headers). All IP devices in same IP subnet (as defined with IP address/netmask pair) must use same MTU. If not, then packets larger than MTU setting will be dropped by receiver (smaller packets are fine). Mind tgat no IP packet fragmentation takes place for communication within same subnet, only routers perform fragmentation (or more often don’t) if ingress packet is larger than MTU on egress interface.
And yes, L2MTU has to be large enough to hold MTU-sized L3 packets.
When talking about bridge MTU, it’s a setting of “the switch-facing interface of the router” (read more about different bridge personalities) … and by “router” in this sentence I’m referring to any ROS-running device with bridge and IP address set on bridge … can be switch with management interface as well (I copy-pasted the phrase used in the linked document).
L2MTU is maximum size of ethernet frame that switches (and L3 devices) can receive, store and forward. For switches/bridges it’s more or less configuration of available memory for buffering frames, for L3 interfaces (e.g. NICs) as well. So in principle L2MTU could be set to an arbitrary large value (OK, most devices have upper HW-limit), but that would likely limit device throughput due to reduced number of buffered frames (number of frames times maximum expected/allowed frame size must me dmalker or equal to available memory size).
And it’s sensible to have L2MTU (roughly) equal on all bridge/switch ports, you don’t want bridge/switch to drop large frames because they don’t fit egress port’s (small) L2MTU.
Even if you don’t foresee some port to carry jumbo frames for unicast communication … there are broadcast packets which by their nature are supposed to reach each and every device in broadcast domain (that’s L2 equivalent of an IP subnet).