feat: reco training script with multi paths sets

[charlesmindee]

This PR adds the option to train recognition with multiple datasets.
the arg data_path in the script is replaced by a train_data_path, a list of 1 or more paths and val_data_path, a path.
The structure inside each path must be: path/images/ and path/labels.json.
Any feedback is welcome!

[codecov]

Codecov Report

Merging #377 (982a1e6) into main (8e14332) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##             main     #377   +/-   ##
=======================================
  Coverage   96.25%   96.25%           
=======================================
  Files          83       83           
  Lines        3554     3554           
=======================================
  Hits         3421     3421           
  Misses        133      133           

Flag	Coverage Δ
unittests	96.25% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8e14332...982a1e6. Read the comment docs.

[fg-mindee]

Thanks for the PR! I'm not sure passing a list to a bash script is a good idea. Say we have dozens of parts, it would be better to assume that all parts are the subfolders of train_set for instance (train_set/part1, train_set/part2, etc.). If the json labels are directly in this folder, then there is only 1 part, if not, we instantiate the set with the first part, then merge the others. What do you think?

[fg-mindee]

One last improvement suggestion and we're good to merge!

[fg-mindee]

Thanks again!

[fg-mindee]

Thanks!