sync w/ internal repo; update models

[leewyang]

This PR syncs the migrated QualX code with the latest code from the internal repo.

Changes

1. Fix failed stages handling.
2. Ignore wholestagecodegen for unsupported.
3. Use actual gpu app duration.
4. Default to dMAPE metric.
5. Add updated model json files.

Test

Following CMDs have been tested:

Internal Usage:

python qualx_main.py preprocess
python qualx_main.py train
python qualx_main.py evaluate

[parthosa]

Thanks @leewyang. LGTM. Welcome to tools repo 🎉

Once things are stable, It might be helpful to work on a separate feature branch to skip GitHub workflows related to maven.

cc: @amahussein

[amahussein]

Thanks @leewyang
A couple of issues I raised in my comment.

[amahussein]

Lets set the standard:

• raising an error: it means something is really bad. In that case do we fallback to legacy Speedup?
• Return a default DF with speedup 1.0: this is the case when the QualX says that there is no speedup because there are ceratin features, files/columns I cannot read/empty. Then the speedup is 1.0

For the case when it is "not toc_list", do we want to raise an error? or return 1.0 after logging that files were missing?
CC: @eordentlich and @parthosa

[leewyang]

@amahussein this should be a very rare case. Basically, the user has requested to load CSV files for a dataset, and we have found none at all for the entire dataset. Previously, we would just skip this dataset (failing silently), so added this code to raise it to the user's attention. Presumably, the legacy Speedup would also fail in this case too (or would it just return 1.0?).

[amahussein]

I thought that change was rolled back..Just double checking.

[leewyang]

Per @eordentlich's commit to internal repo, this is a workaround for #860.

[amahussein]

Let's merge that to unblock then followup later on whether to raise-error or not in subsequent changes.