Pykelihood

New features

* The `Distribution.fit` method accepts a `scipy_args` dictionary which is
passed to `scipy`'s `minimize` function.
* The confidence interval computed by the profiler now uses root finding to
find the bounds where the likelihood ratio test starts failing. This means
confidence intervals can only be computed for the distribution's parameters.
* Upper bounds on dependencies were removed, improving compatibility with
recent versions.

Breaking changes

* Renamed `stats_utils` module to `profiler`
* Data must now be provided to kernels on creation, unbound kernels are
no longer allowed
* `Parameter`s are no longer subclasses of `float`, use `.value` to get
their stored value
* `ConditioningMethod`s were removed, their uses can be replaced with
_score functions_
* The `biv` parameter to the `Profiler` was removed, confidence
intervals are univariate only

Removed

Many distributions and utilities which were created with a specific use
case in mind and aren't generally useful have been removed:

* `MixtureExponentialModel`,
* `ExtendedGPD`,
* `PointProcess`,
* `CompositionDistribution`,
* `DetrendedFluctuationAnalysis`,
* `pettitt_test`,
* `threshold_selection_GoF` and `threshold_selection_gpd_NorthorpColeman`,
* extreme values visualisation routines,
* process samplers (Poisson and Hawkes).

New features

* Metrics: `{pp,qq}_l{1,2}_distance`, `likelihood`, `expo_ratio`
* Log-normal distribution
* Plotting functions now accept an `ax` argument to use instead of the
global `plt` figure
* Constant kernel (most useful for testing)
* `Kernel`s have a `with_covariate` method that returns a new kernel
with the provided data as covariate, but all parameters are kept the
same
* The `random_state` parameter to the `Distribution.rvs` method is now
explicit and no longer hidden in the `**kwargs`

Bug fixes

* Fixed `fit_instance` for nested kernels with fixed values
* Fixed the `TruncatedDistribution` which forgot its bounds after fitting
* A parameter which shows up in several places in a distribution will
keep the same value when fitting instead of returning independent
parameters

Other

* Add section to README on fitting other score functions than the likelihood
* Add changelog with all version changes up to this one

Various bug fixes due to new names in version 0.3.0

New features

* New trigonometric kernel

This release aims at making fitting even more generic by replacing
penalized likelihoods with score functions

Breaking changes

* Remove `log_likelihood` and `opposite_log_likelihood` methods from
`Distribution`s
* Remove `penalty` argument from `fit`
* `ConditioningMethod`s take `distribution` as the first argument and
`data` as the second
* Rename `Likelihood` class to `Profiler`
* Rename `mle` attribute to `optimum` in `Profiler` class

New features

* `fit*` methods minimize a `score` function which takes a distribution
and data as arguments. By default, they maximize the log likelihood.
* New `metrics` module which contains scoring functions
* The `Profiler` has a `score_function` argument
* Add `threshold_selection` based on a multiple threshold penultimate model
* New QQ-plot visualization methods

Bug fixes
* `Distribution.rvs` now passes all parameters to `scipy`'s `rvs` methods.
In particular, `random_state` is now propagated.
* `TruncatedDistribution` now ignores data outside its range

Other

* Improve overall typing of the library

New features

* Add `TruncatedDistribution`

Bug fixes

* Avoid replacing non-constant parameters with `ConstantParameter`s
inside `ParametrizedFunction`s
* Fix nested kernel fitting