[dask] [python] Store co-local data parts as dicts instead of lists

[ffineis]

Addresses #3795. The _train function in dask.py places delayed "parts," i.e. data, label, sample_weight, and group data into lists (which get distributed to workers) to enforce co-locality of data. Each worker is distributed a list_of_parts, a list of lists. Once the worker receives its lists of parts, it infers the data, label, and then optionally sample_weight and group from each list_of_parts by assuming that data is in position 0, label is in position 1, and if the model is not a DaskLGBMRanker, the second position is reserved for sample_weight.

This is an issue of readability + scalability - implying the presence of sample_weight and group based on the length of each sublist of list_of_parts in conjunction with the estimator type is... hard to understand and doesn't lend itself well to storing additional parts like eval_sets. Instead, let's just store individual sets of parts as dicts instead of lists. Now each worker will receive its portion of the overall dataset as a list of dicts instead of a list of lists.

[jameslamb]

This is great, thank you! I just have one very small nitpicky style thing.

[jameslamb]

This is much cleaner and I think it will help reduce the risk of mistakes in the future. Thanks so much!

[StrikerRUS]

Awesome! Thanks! Code is much cleaner now.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.