:= is a left-over assignment symbol from earlier R/S days. Be default it currently has no implementation. Packages such as data.table, and wrapr define their own versions, but these are not available until you call library(pkg). I think dplyr just uses the presence of the symbol in unparsed expressions (so does not need an actual implementation) and currently does not export :=.
I think you can get around the check with something like the following (just loose in your package, and don't bother to export it).
`:=` <- function(a,b) {stop(":= should not be called directly")}
Then code such as the following still works (as dplyr either uses its own package definition, or more likely never executes the user code in the first place).
dplyr::mutate(data.frame(x = 1:2), y := x + 1)
The above is safe code, and turns off a minimum number of checks.
By the way wrapr defines := as assigning names, which is very close to how a user sees := being used in other packages. For example.
rm(list = ":=")
library("wrapr")
c("name1", "name2") := c(1, 2)
# name1 name2
# 1 2