[ENH] Improve performance of TimeSeriesDataSet.__getitem__

[denix56]

Description

Pandas DataFrame is quite slow in comparison to numpy due to additional checks.
By replacing it with np.recarray I was able to improve performance by 5-10%.
Recarray allows us to have nice attribute access as in pandas, while improving performance.
The raw numpy arrays are a bit faster than recarray, however the difference is not as big as between pandas and recarray.
I have tested on Demand Forecasting with gpu=1, 0 workers and pin_memory=True.

[codecov-commenter]

Codecov Report

Merging #806 (eb706f9) into master (0b5892a) will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #806   +/-   ##
=======================================
  Coverage   89.05%   89.06%           
=======================================
  Files          24       24           
  Lines        3829     3832    +3     
=======================================
+ Hits         3410     3413    +3     
  Misses        419      419           

Flag	Coverage Δ
cpu	89.06% <100.00%> (+<0.01%)	⬆️
pytest	89.06% <100.00%> (+<0.01%)	⬆️

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

Impacted Files	Coverage Δ
pytorch_forecasting/data/timeseries.py	93.12% <100.00%> (+0.02%)	⬆️

---

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 0b5892a...eb706f9. Read the comment docs.

[jdb78]

I am tempted to merge this. Think we should run the example notebooks also because things might change there - even if only visual.

[jobs-git]

any news on this?

[fkiraly]

I merged this manually, as I had already edited the same location in __getitem__, and the file has moved.

How would we know this is an actual performance improvement? Have you tested it, @jobs-git?

Let's see if the tests pass.

[fkiraly]

It appears the changes in this PR break internal API assumptions in other methods, e.g., get_groups - so it cannot be merged in its current state.

Still worth to keep open as long as we are reworking for v2.