Fix612 lazy download dataset

[PGijsbers]

Datasets can now be downloaded without downloading the arff file.
Function signature of get_dataset changed from
def get_dataset(dataset_id: Union[int, str]) -> OpenMLDataset:
to
def get_dataset(dataset_id: Union[int, str], download_data: bool = True) -> OpenMLDataset:
I chose to default to True so there are no breaking changes to existing code.
If download_data=False, only metadata will be downloaded (i.e. all data except the arff file).

Whenever a user invokes retrieve_class_labels or get_data, both of which require the arff-file, the arff-file is retrieved and processed as if it were downloaded from the server on initialization (e.g. pickled). This happens without warning or error/argument.
As long as only functionality is used which does not require the arff file, no additional data is downloaded.

I took the liberty to refactor retrieve_class_labels s.t. it uses the already downloaded feature metadata instead of reading the arff file. This makes it so that retrieve_class_labels can 50used without downloading the underlying data, and overall should really speed up the method in cases where a huge arff file was loaded just to check the header 👍
As part of this I also addressed open issue #507 that was being worked on in #508 (I only realized afterwards). Though it looks like that code was from before @janvanrijn changed the xml so that the nominal values are a list. My solution is able to use this update and with only minor changes address the issue.

Fixes:
#643
#612
#507
#446
#346 in part

[codecov-io]

Codecov Report

Merging #644 into develop will increase coverage by 0.54%.
The diff coverage is 92.62%.

@@             Coverage Diff             @@
##           develop     #644      +/-   ##
===========================================
+ Coverage    90.13%   90.67%   +0.54%     
===========================================
  Files           32       32              
  Lines         3366     3390      +24     
===========================================
+ Hits          3034     3074      +40     
+ Misses         332      316      -16

Impacted Files	Coverage Δ
openml/datasets/dataset.py	88.41% <90.76%> (+6%)	⬆️
openml/datasets/functions.py	92.04% <94.59%> (-0.7%)	⬇️
openml/utils.py	92.06% <95%> (+0.55%)	⬆️

---

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 94102f3...b812904. Read the comment docs.

[PGijsbers]

It looks like continuous-integration/appveyor/pr failed because the conda update stalled.
I can't figure out how to rerun it, without pushing a new commit to this PR. I tried to re-build the commit, but that only seems to trigger continuous-integration/appveyor/branch (which worked but is now being rerun).

[PGijsbers]

I can't seem to restart the continuous-integration/appveyor/pr build tests without submitting a new commit. If requested, I will add one, but I feel confident that it is not needed to have said build rerun and the code quality is fine (as all three other builds passed).

[joaquinvanschoren]

Awesome! ::hooray::