Matplotlib-inline-labels

This realase brings:
- Bug fixes
- Test enhancements

Bug fixes
---
- Fixed debug figure display on interactive display backends (tested for tkAgg)
- Fixed imports from different modules, which were broken since v0.2.1 when code was splitted into several modules

This realase brings:
- Bug fixes
- Test enhancements

Algorithm
---
- **Post-processing** stage has an enhanced choice in the default algorithm.
- When looking for the longest group of position and rotation adjacent candidates, instead of choosing the first longest group based on cardinality, now we choose among groups of highest cardinality the one with the longest distance on line chunk measured between its extremities. In the improbable case where there would still be several (e.g. highest cardinality groups all with only one candidate), the first one is chosen.

Two dedicated data structures `Labels_lcs_adjPRcs_groups` and `SepLongestSubgroups` have been introduced for the latter purpose, and `Labels_lcPRcs` has been removed
- **Trailing TODO**
- **Processing**: Optimize the function giving the label's bounding box rotated and translated geometry, and its usage
- **Processing and post-processing**: use of more discrete separation levels or a move to continuous separation levels
- **Post-processing**: use multi-objective optimization based on pymoo to enhance labels' position among themselves
- **Post-processing**: handle the case where two labels overlap because the highest separation level found, is 0 on the two corresponding lines. This will be handled after multi-objective optimization implementation


Misc
---
- Added the usage of `pytest-xdist` to speed up testing in GitHub actions CI, through parallelization
- Enhanced test coverage

Bug fixes
---
- Fixed a bug in line geometric modeling at the closed curve detection stage: closed curves with hidden tails (`np.nan` values) were not properly detected
- Fixed a bug in publish to pypi Github action: signing the dists with Sigstore was not working since `sigstore/gh-action-sigstore-python` version was not fully specified
- Switched back to absolute links in `README.md` to have example images properly displayed on pypi project page (other branches have their `README.md` file pointing to the wrong images, those of the main branch). It's still not ideal
- Fixed a bug on labels' bounding box dimensions on visual debug figure. Dimension of the last label processed was used for all labels instead of using each label's bounding box dimension
- Fixed a bug preventing to show visual debug figure legend when all labels could be inlined
- Fixed the minimum number of label's center position candidates to be 3 instead of 4
- Fixed the line chunks polygonization at the geometry extraction stage, by applying the polygonization to the union of all continuous non `nan` data segments instead of to each one separately

This release brings:
- performance enhancement, 20% to 50% faster than previous release
- a better handling of closed curves

Algorithm
---
- **Pre-processing** stage of creating label position candidates has been enhanced with:
- a safeguard on the number of candidates via `maxpos` arg preventing cases of low `ppf`, or small label `fontsize` or high figure `figsize` or a conjunction of all, to lead to too many unnecessary position candidates
- handling of line self-intersections
- better handling of closed curved (position candidates added near the junction)
- filtering of position candidates too close to any "distant" part of the same line (folds) or to any other line of the plot
- filtering of position candidates on a line part with a too high curvature for placing the label properly. Two performance equivalent options are available in the code `fast` and `precise` curvature estimation near the line chunks' edges. `fast` is used as default
*The gain on faster pre-processing is lost on processing, and the additional time used by the precise curvature estimation is compensated by faster processing*
- **Processing** stage of finding the best label rotation for each label position candidate has been enhanced with:
- limitation of the rotation tested angles, with the label's bounding box geometry, to pre-estimated best rotation +/- 2 x arctan(l/L)
- limitation of the label separation computing to the label rotation with the smallest alignment error
- creation of a flat list of position and rotation candidates structure for use in future post-processing algorithms

The function giving the label's bounding box rotated and translated geometry, and its usage could still be optimized

The use of more discrete separation levels or a move to continuous separation levels still has to be explored to see if it leads to better label inline placement
- **Post-processing** stage has a draft algorithm using multi-objective optimization based on pymoo to enhance labels' position among themselves

Code structure
---
- Data structures have been refactored for better readability and put in a dedicated file `datatypes.py`
- Functions of the main file `inline_labels.py` have been splitted in:
- **drawing** (`drawing.py`): retrieval of matplotlib objects and creation of the visual debug figure
- **geometries** (`geometries.py`): conversion of matplotlib objects to geometric objects
- **pre-processing** (`preprocessing.py`): identification of label placement candidates on lines and filtering with line intersections and high curvatures
- **processing** (`processing.py`): identification of label rotation and separation
- **post-processing** (`postprocessing.py`): selection of the best candidate for each line

Utils
---
- Timers are available to show the cumulated time of functions. Timers can be set as decorators (to be added or uncommented) on a function that could be optimized
- A timer decorator is available for the main function `add_inline_labels` for overall performance debug. It has to be uncommented to be used

Misc
---
- Code has been reformatted with a max line length of 92 for better readability
- Github actions have been upgraded to last currently available versions
- Dropped python 3.11 compatibility to be able to use hashable slices. Maybe retintroduced later if requested, provided more advanced multi-objective optimization is used in post-processing
- Added a step in publish to pypi Github action to extract release notes from `NEWS.md`