Operator[] on functions confusing

[gchabert]

Currently, if f is a n_m matrix-valued function, the components of f are the n_m entries.
We would expect f[0] to be the first row, not the entry (0,0).

The code is easy to change but we may lose memory optimization. E.g, building the function f:x -> eye(10) will actually build only 11 functions (the diagonal entries + one null function) whereas introducing each row function will lead to 30 functions (one for each row, one for each "zero" function inside the row, and one for the diagonal element).
In the other hand, we also improve efficiency when one has to evaluate a row simultaneously.