Proposal to migrate `DelayedMask` to the DelayedArray package

[LTLA]

My alabaster.matrix package has a DelayedMask class that replaces all instances of a placeholder with an NA value during extract_*_array() calls. Essentially, it is equivalent to:

library(DelayedArray)
y <- DelayedArray(matrix(rbinom(100, 1, 0.2), ncol=5))
placeholder <- 1

z <- y
z[z == placeholder] <- NA

# Gives the same (abstract) matrix contents as:
alabaster.matrix::DelayedMask(y, placeholder)

AFAICT the DelayedMask approach should be more efficient as the DelayedMask only realizes one matrix while the delayed [<- requires realization of both z and z == placeholder. (Currently DelayedMask is literally for masking NAs but it could be generalized to handle more boolean operations, e.g., so z[z > cap] <- cap could be used to upper-bound the contents of the matrix.)

Would this be of interest to the DelayedArray package? Perhaps there might be an easy way that we can detect that the logical subset is generated from the same matrix that it is operating on in [<-, and switch from creating a DelayedNaryIsoOp to a DelayedMask.