Julia shouldn't crash but rather raise an error about the existence
of a function generated by metaprogramming. I didn't manage to isolate
the segfault withtout the package.
To overcome the issue, I add another level of multiple dispatch/abstract
type for synchronized models.
Test of the euclidean distance automaton works.

This commit groups the change operated to the creation of models and
simulate function of a ContinuousTimeModel.
The general idea is to create a concrete type and a simulate function
 per model creation by metaprogramming.
- Now, ContinuousTimeModel is an abstract type. Each creation of a model
defines a concrete type T <: ContinuousTimeModel by meta programming.
- f! and isabsorbing ContinuousTimeModel fields are Symbols.
- simulate(::ContinuousTimeModel) is run by multiple dispatch, according
to the type of the model.

Can't run the whole tests for now but unit/simulate_available_models.jl
runs properly (i've updated the list of models in this commit), and I've
manually checked in the repl that simulations run correctly (distributed
/ plots).

check_constraint and update_state functions are created inside
create_automaton_*(), by the evaluation of meta_elementary_functions <:
Expr. check_constraint and update_state don't depend on constants::Dict
anymore: constants are interpolated inside the expression.

Add of tests for euclidean distance automaton.
Add a new function euclidean_distance that computes the euclidean
distance of a Trajectory compared to observations along a discrete
timeline.

automata and proba moels are differents. This bug was introduced after
we change the structure of automata in order to better distributed
programming.
It leads to a restructuration of automata. In fine, the structure is
more stable and consistent. It is possible it is more computationally
efficient, but for now it is not well tested.

2) Creation of functions that collects data along with a timeline
over a trajectory + unit test vectorize.

3) Very small fix in plots and abc smc.

models. It's a better semantic and improve performance as well as
readability of the code. All the tests passes (except the remark in the
last commit).

copyto! is implemented for LHA states.

Automaton G works almost well, a bias in the distance is identified
compared to Cosmos when one of the model parameters is zero.

Implementation of volatile_simulate, which simulates SynchronizedModel
without saving values: it only returns the last LHA state of the
simulation.

Implementation of distributed computations of variable automata with
volatile_simulate.

Tests passed.

Fix about methods of SynchronizedTrajectory + tests.

Simulation for ParametricModel works.

small.

Fix of a misbehaviour of isbounded() for trajectories.

order to replace vcat by push!.

The benchmarks are at least equal, and better when the buffer size is
small. However, no significative difference of performance with adequate
buffer size.

My opinion the gain can be seen when simulations are going to be very
long with more complicated models than we have implemented for now.

All tests passed.

Add of a new type of Trajectory: SynchronizedTrajectory which is created
by the simulation of a synchronized model with an LHA.

Simulate function for SynchronizedModel executes and is tested by two
simple models.

Next step : Cosmos based tests.

an object.

Type stability of simulate, dist, read_trajectory seems to be preserved.

First implementation of LHA.
Synchronized simulation not yet tested.
Reading of trajectories with LHA seems to work with one test basted on
simulations that checks that the last value is well read.

Major fix of type unstabilities which improved performances.
For now on, we observe for pygmalion benchmarks:
+2x speed simulation compared to simulation of models.
100x speed in terms of reading values (no allocations).

Add of bench for @view macro.

Fix of a major issue in the simulation: sometimes the number of states
wasn't consistent. Add of test to check the consistency of the
Trajectory structure (unit/check_trajectory_consistency).

Implementation of the lp distance for trajectories.
Add of unit tests for lp distance.

can be set by the user. Changes of SIR and ER.

previously loaded model). The package warns when this happen.

Several benchmarks for the data structure. Now we chose the row buffer
method according to results (mainly because the trajectory access with
far way faster, but strangly not that much differences in other cases).

- Real => Float64 for better performance (Real is an abstract type)
- Should not create a type inside the model because the compiler fails
at estimating the type of the output of the f function (detected with
@code_warntype)
- Intentionaly ER is not as well improved as SIR in order to check the
differences of performance.

- Add of benchmark scripts that compares perf wrt pygmalion.
=> As expected it is bad in terms of reading cost (because of row by row
matrix read) but it is also not good in terms of simulation cost.
hcat seems to perform badly, should investigate.

SIR model.

- Evolution of the ContinuousTimeModel type
- simulate, easy access to Trajectory fields
- fix of several errors

Two folders:
- core/ which contains the essential files for the packages
- tests/ which contains tests and benchmarks of other packages / methods

For now on:
- we wrote minimal tests in tests/ (a simulation of SIR model) that should run for
a first minimal version of the package
- we described in core/ the general structure of essential
types, how they articulate each others and the minimum methods we should
implement for a first version that runs the tests.