Add na.rm to aggregating functions

[TuomasBorman]

This PR adds na.rm option to following functions: sumCountsAcrossFeatures, numDetectedAcrossFeatures, summarizeAssayByGroup, numDetectedAcrossCells.

This PR is related to this issue: #33

[LTLA]

Thanks for the PR. I'll address some immediate points first, but having slept on your problem, I would recommend a much simpler solution.

• value == NA_INTEGER is not correct if the matrix was double-precision, as the NA_INTEGER value (-2^31) is a valid double-precision number. You'll need to check the input matrix type and only perform that check if the input matrix is integer or logical.
• Conversely, std::isnan() is unnecessary when the input matrix is integer or logical, as IEEE specials wouldn't fit inside a 32-bit integer anyway.
• All of the new checks are unnecessary when na.rm=FALSE. To avoid having a potential optimization-breaking if statement in the tight inner while loop, hoist it out and duplicate the for (auto s : runs) loop for each NA'ness status (na.rm=TRUE with double, na.rm=TRUE with int, na.rm=FALSE) with only the relevant NA checks in each one.
• You have too many comments - don't touch the lines that you don't need to.
• No pipes in my packages.

You can see that it is quite tedious to protect against NAs while preserving performance for the vast majority of (single-cell) use cases that don't need it. So perhaps you might consider a much simpler approach:

y <- rsparsematrix(10000, 5, density=0.1)
z <- sample(LETTERS[1:5], nrow(y), replace=TRUE)
y[1] <- NA

y <- DelayedArray(y)
is.okay <- !is.na(y)
type(is.okay) <- "double"
DelayedArray::rowsum(y, z, na.rm=TRUE) / DelayedArray::rowsum(is.okay, z)

4 lines and done. No copying involved as the DelayedArray() defers the is.na and type coercion operations. Maybe a bit slower than the C++ code but I doubt it matters much.

[TuomasBorman]

Thank you for the feedback! I'm not very experienced with C++, so I jumped straight into the deep end. Your suggestions seems to be working and it is much simpler. I would still tend to think that scuttle is the right place to fix this na.rm issue. Would it be ok, if I

1. revert changes affecting the C++ code
2. add na.rm to .sum_across_features: if na.rm, then run your 4 lines instead of C++ code

Moreover, summarizeAssayByGroup used loops (lapply) to aggregate data. I guess this approach could also improve the efficiency of that function.