Implement preconditionners as in UMFITER and ITERSTACK:
– For UMFITER, on first iteration, do a full LU decomposition, and use it as a preconditionner
  in further iterations (similar to UMFITER in TROLL).
– For ITERSTACK, build a preconditionner based on small LU decomposition
  repeated several times across the stack

By the way, uniformize the preconditioners for solve_algo={7,8} across the
different implementations.

TODO:
– make sure we have a good default for options_.simul.iter_tol. The gmres and bicgstab
  functions default to 1e-6. The run_block_bytecode_tests needs 1e-7. And
  GIMF/GMMET need 1e-12 in the block-decomposed version (partly because tolf=1e-8
  there)
– add tests for the three preconditioners (and ideally their suboptions)
– documentation

 Perfect foresight with controlled paths: no longer crash with homotopy_exclude_varexo option and multiple controlled variables

Most mathematical functions return NaN, ±Inf or a complex number when called
outside of their domain of definition, but erf, erfc, normcdf and normpdf can
throw exceptions when given a complex number (under MATLAB; not under Octave
for erf and erfc).

Treat those cases as a failure to solve for the current homotopy step, and
continue the homotopy procedure, instead of failing outright.

This is useful when there are numerical errors in some files, such as
evaluating erf() on a complex number (under MATLAB this throws an exception;
under Octave, it instead returns a complex number).

Forward both the error message and error ID (but not other properties of the
MException object, there does not seem to be a way to do that from a MEX file).

Immediately deallocate the temporary C string instead of waiting for the MEX to
return.

Ref. !2357

– fix linearity check for Ramsey models with lead/lag>1 on endos or >0 on exos
  (Closes: #1960)
– allow back “residual” as symbol name
– misc cleanup of tokens associated with eponymous std::string

The performance gains were not consistent and were ranging from −10% to +10%
depending on the size of the involved matrices. We also met some numerical
instability problems on one of runners (those without AVX).

 identification: shut off slice_initialize_with_mode option with slice to prevent crash

Closes #1957

See merge request !2356

Closes Dynare/dynare#1957

– gracefully error out when a PAC model is not used in a pac_expectation
  operator
– fix handling of model local variable in DataTree copy constructor

[skip ci]

In particular, this now removes float.h which was conflicting with the one from
a recent MinGW snapshot.

By the way, rename “stop” variable to “converged” for clarity.

Ref. !2353

[CI skip]

 OccBin: disable incompatible options

See merge request !2352

Ref. !2351

Perfect foresight: improve sanity checks for linear_approximation and endogenous_terminal_period options

 Fix crash of perfect foresight w/ expectation errors + marginal linearization + several learnt_in blocks

[skip ci]

Ref. !2350

Prevents cryptic crashes

Leads to a simpler code.

I am not sure I want to repair this...