Remove all nan to double conversion in infer_feature_types

[freddyaboulton]

Pull Request Description

Fixes #3194

Perf tests:
nan-to-double-reports.zip
New branch seems to be faster on large datasets for search and model understanding (nyc-taxi, regress, kddcup)

---

After creating the pull request: in order to pass the release_notes_updated check you will need to update the "Future Release" section of docs/source/release_notes.rst to include this pull request by adding :pr:123.

[codecov]

Codecov Report

Merging #3196 (d694d91) into main (aa0ec23) will increase coverage by 31.9%.
The diff coverage is 100.0%.

❗ Current head d694d91 differs from pull request most recent head c530d4a. Consider uploading reports for the commit c530d4a to get more accurate results

@@           Coverage Diff            @@
##            main   #3196      +/-   ##
========================================
+ Coverage   67.9%   99.7%   +31.9%     
========================================
  Files        326     326              
  Lines      31392   31131     -261     
========================================
+ Hits       21305   31033    +9728     
+ Misses     10087      98    -9989     

Impacted Files	Coverage Δ
...onents/transformers/imputers/per_column_imputer.py	100.0% <100.0%> (ø)
evalml/pipelines/utils.py	99.4% <100.0%> (+0.1%)	⬆️
...l/tests/component_tests/test_per_column_imputer.py	100.0% <100.0%> (ø)
...valml/tests/component_tests/test_simple_imputer.py	100.0% <100.0%> (ø)
...del_understanding_tests/test_partial_dependence.py	100.0% <100.0%> (+100.0%)	⬆️
evalml/tests/pipeline_tests/test_pipeline_utils.py	99.7% <100.0%> (+1.3%)	⬆️
evalml/tests/utils_tests/test_woodwork_utils.py	100.0% <100.0%> (ø)
evalml/utils/woodwork_utils.py	100.0% <100.0%> (ø)
evalml/data_checks/outliers_data_check.py	98.1% <0.0%> (-1.9%)	⬇️
... and 76 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 aa0ec23...c530d4a. Read the comment docs.

[bchen1116]

This looks good to me! I left a few comments on updating column names.
As for the perf tests, do you have a hypothesis on why some model understanding tests increased a fair bit in time?

Not blocking, but curious as to these slight regressions

[bchen1116]

Why did this need to change? Also, we should rename this column since there are no longer nans

[freddyaboulton]

Sorry I thought I had answered this yesterday.

The only way to pass an all-NaN column to the imputer is if it's not typed as Unknown, which is what the default woodwork inference returns. So we need to init woodwork on the dataframe. However, because of #2055, the imputer will modify the input data.

So the some_nan column will have the NaNs imputed. We can't change the name of the column because the intention is to compare the input data before and after running through the imputer and the column name doesn't change after going through the imputer.

[bchen1116]

same here

[freddyaboulton]

Same comment as above ^

[bchen1116]

rename

[angela97lin]

Looks good, thank you @freddyaboulton!

[angela97lin]

This change is because all_null will be considered an unknown column now, rather than double right?

[freddyaboulton]

Yes!

[freddyaboulton]

@bchen1116 I think the performance on model understanding is faster on this feature branch! See datasets like nyc taxi and regress.csv. The particular dataset you've pointed out is slower but the difference is 0.1 seconds, which I don't think is meaningful.