Allow argument to specfiy how auto-sklearn handles compressing dataset size

[eddiebergman]

• Builds off Dataset size reduction fixed, updated TargetValidator to match signatures #1250 which introduces parametrized dataset compression and uses the custom splitters.
• Still requires documentation, should probably wait until use_new_splitter is merged so the new updated docs can be written to.

Add an argument to AutoSklearnEstimator that allows the user to specify how we perform dataset reduction and whether it should be enabled or not. This also serves as some documentation as to how we perform dataset compression, such as precision reduction, subsampling and what issues may arise.

AutoSklearnClassifier(
    ...,
    dataset_compression: Union[bool, Dict[str, Any]] = True
)

• Can enable disable with True or `False
• Can pass in a dict (implying True)
  • "memory_allocation" - A float for a percentage of memory limit or int for absolute memory
  • "methods" - List of of dataset reduction to perform
    • "precision" - perform float precision reductionn
    • "subsample" - subsample the data so that it fits into the memory allocation.

# default
{
    "memory_allocation": 0.1,  # How much of memory to compress the data into
    "methods": ["precision", "subsample"],  # Which methods to use
}

A user can specify only some of these keys or pass None and we fill the rest with these default values. Includes validation of the passed items.

Also updated the docstring for resampling_strategy to inform users that they may wish to disable this if using custom object splitters.

These PR addressess:

• Getting only TIMEOUT for PredefinedSplit #1274 - User trying to use predefined split, would fail if the dataset gets reduced.
• memory_limit interferes with "resampling_strategy=GroupKFold" #1137 - The groups specified no longer align with the reduced dataset size.
• User has no way to control dataset size reduction for large datasets and small memory #1279 - Directly mentions what this PR addresses

[codecov]

Codecov Report

Merging #1341 (f03f76e) into use_new_splitter (42e4397) will decrease coverage by 0.42%.
The diff coverage is 100.00%.

@@                 Coverage Diff                  @@
##           use_new_splitter    #1341      +/-   ##
====================================================
- Coverage             88.82%   88.40%   -0.43%     
====================================================
  Files                   140      140              
  Lines                 11912    11245     -667     
====================================================
- Hits                  10581     9941     -640     
+ Misses                 1331     1304      -27     

Impacted Files	Coverage Δ
autosklearn/automl.py	86.98% <100.00%> (+0.11%)	⬆️
autosklearn/estimators.py	93.45% <100.00%> (+0.03%)	⬆️
autosklearn/experimental/askl2.py	83.33% <100.00%> (ø)
autosklearn/util/data.py	82.08% <100.00%> (+4.73%)	⬆️
...sklearn/pipeline/components/regression/__init__.py	83.52% <0.00%> (-6.31%)	⬇️
...arn/pipeline/components/classification/__init__.py	84.94% <0.00%> (-5.90%)	⬇️
autosklearn/pipeline/components/base.py	78.78% <0.00%> (-3.33%)	⬇️
autosklearn/evaluation/__init__.py	85.47% <0.00%> (-2.91%)	⬇️
...ata_preprocessing/categorical_encoding/__init__.py	85.71% <0.00%> (-2.80%)	⬇️
...eline/components/feature_preprocessing/__init__.py	89.33% <0.00%> (-2.26%)	⬇️
... and 8 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 42e4397...f03f76e. Read the comment docs.