support feature or sample wise missing values

[rmflight]

Currently only a single missing value is supported data-set wide, by providing a single zero_value.

It should be possible to provide feature or sample wise zero values by the user providing a vector that is the same length as either the columns or rows, this would determine whether it is feature or sample wise.

[rmflight]

The way to implement this is at the R code level, where the imputed values are replaced with NA in the data matrix, and then the C++ code handles the missingness.

[rmflight]

After discussion with @hunter-moseley, a design would be to create new parameters:

global_na = c(NA, Inf, 0)
feature_na = NULL
sample_na = NULL

So then if the user changes the global_na they can do that.