Seems like a bad idea to coerce a DelayedMatrix to the seed class

[LTLA]

I noticed that colVars (and possibly other functions) has this little nugget:

DelayedMatrixStats/R/colVars.R

Lines 71 to 74 in 670e46e

	message2("Coercing to seed class", get_verbose())
	# TODO: do_transpose trick
	simple_seed_x <- try(from_DelayedArray_to_simple_seed_class(x),
	silent = TRUE)

I'm not sure this is a good idea; does this respect the block size and memory constraints of block processing? For example, if I defined something like this:

library(DelayedArray) 
y <- DelayedArray(rsparsematrix(10000, 10000, density=0.1)) + 1
 
library(DelayedMatrixStats) 
library(sparseMatrixStats) 
colVars(y)

If seedClass() was smarter than it currently is and detected that the dgCMatrix at the core of y now has a colVars() method, the function would attempt to realize the dense matrix in the above example - not good. Fortunately, seedClass() is not very smart and returns DelayedUnaryIsoOpStack, which means that we always go through block processing. This means that either seedClass()'s lack of intelligence is intentional, in which case it seems like isPristine() will always be true at that point in the function and there is no point in the else expression; or it is meant to be more intelligent, in which case memory constraints are likely to be violated.

Why not just have a really simple rule:

    if (isPristine(x)) {
        candidate <- selectMethod("colVars", seed(x), optional=TRUE)
        if (!is.null(candidate)) {
            return(candidate(seed(x), ...)) # fill in the other arguments...
        }
    }

    # block processing here.