(manually cherry picked from commit c17bbe19)

(manually cherry picked from commit 7743fc6f)

(manually cherry picked from commit 11e9e608)

(cherry picked from commit 4832c0dd)

It was no more possible to fill a variable with NaNs, after the switch
to the new behaviour of the merge method (see test #5 in
@dseries/merge.m). Added a legacy mode to the merge method to
circumvent the issue (@dseries/subsasgn sets legacy equal to true when
calling the merge method).

(cherry picked from commit 67077fe8)

Second argument of merge overwrites its first argument except if
second argument has NaNs (in this case first argument takes
precedence). This commit reverts fd7476e6 and adds a new unit test.

(cherry picked from commit c5b190d8)

Normally it should still be possible to instantiate dseries objects
from mat files generated by previous versions of Dynare (before 4.6).

warning doesn't print sprintf characters unless it is given more than one argument. hence use sprintf

If a variable, say x, is defined both in dseries objects o and p, then
merge(o, p) will select the variable x as defined in the second input, p,
except for NaNs elements in p if corresponding elements in the first
input, o, are numbers.

— On GNU/Linux, the locally installed binary now has priority over the
  system-wide binary
— Do not require 32-bit binary if we are on a 64-bit system and the 64-bit
  binary is there (or vice versa)
— Code factorization in the detection logic

Closes: #41

On Windows and macOS, those binaries are now bundled in packages. On Debian,
one can use the official x13as package.

All issues are now fixed, and we want to detect regressions early.

Also shorten test names, because “testsuite_old_matlab” is too long and gets
truncated on the web interface.

For unknown some reason, this fixes the test under Octave when run from the
whole testsuite.

Previously, .mdl files would unnecessarily be passed through importdata(),
which crashes on Octave.

The nested “if”s (testing for the output file extension) have been reorganized,
so that importdata() is not called on .mdl files (and neither on .acm and .rcm,
where is not needed either).

@rdivide is actually more correct, because we’re not doing an
element-by-element division.

And, more importantly, MATLAB R2009b does not accept @mrdivide in bsxfun.

Provide alternative implementations, taken from Dynare.

prod() does not accept logical arrays, need to explicitly convert to doubles.

MATLAB R2009b is not able to handle HTTPS requests.

Also, having the data in the git makes the testsuite robust to network failures.

Provide alternative implementations, taken from Dynare.