Move default dataset creation responsibilities from runners to the KedroDataCatalog

[ElenaKhaustova]

Description

The suggestion is to move default dataset creation responsibilities from runners to the KedroDataCatalog. This is part of the runners refactoring to decouple runners and catalog.

Context

Currently, we set the default dataset patterns for each runner separately:

kedro/kedro/runner/sequential_runner.py

Line 44 in 06d5a69

default_dataset_pattern = {"{default}": {"type": "MemoryDataset"}}

,

kedro/kedro/runner/thread_runner.py

Line 56 in 06d5a69

default_dataset_pattern = {"{default}": {"type": "MemoryDataset"}}

Then in the AbstractRunner.run() we add these patterns to the catalog before the run and remove them after execution, So at the execution time all intermediate datasets not set in the catalog explicitly are treated as MemoryDatasets.

kedro/kedro/runner/runner.py

Line 114 in 06d5a69

catalog = catalog.shallow_copy(

We can't just add this pattern as the catalog default, as there will be no mechanism to differentiate what dataset is actually in the catalog. Another difficulty is that the catalog is not aware of how it is used by the external objects—it doesn't differ the runtime from other usage. Adding the last will still keep runners-catalog coupling but move it to the catalog side.

Some exploration is needed first to decide how the above problems can be solved at once.