[Rd] patch to model matrix to improve error messages

[Ben Bolker]

   I posted this to r-bugzilla already 
<https://bugs.r-project.org/show_bug.cgi?id=18860>, but I would be 
interested to see if anyone here has comments (fewer people track 
r-bugzilla closely ...)

   The issue: if a variable reference in a model.frame() formula is 
missing from the data but matches a function found elsewhere in the 
environment (e.g. `count()` in `dplyr`), the  user gets the opaque error 
message "object is not a matrix".

   If the variable reference is missing from the data, but matches (e.g. 
a list or list-like object (e.g., a data frame), then we get the more 
useful error "invalid type (TYPE) for variable 'VARIABLE'"

   The issue is that model frame tries to compute the number of rows of 
the data before it tests the type of each column. If we switch the order 
of operations so that the columns are tested first, we get a much more 
useful error message.

   This change seems harmless, but of course I wouldn't be surprised if 
someone can come up with a reasonable scenario where it causes problems ...

    More details/examples in the bug report linked above.  The relevant 
bits of source code are here

model.c row/column checking:

https://github.com/r-devel/r-svn/blob/63369bfed9330b9461ec1d8b90d7251c0118508f/src/library/stats/src/model.c#L138-L152

nrows() function:

https://github.com/r-devel/r-svn/blob/63369bfed9330b9461ec1d8b90d7251c0118508f/src/main/util.c#L81-L94

   cheers
    Ben Bolker