bulk_register can fail when existing trials are all Failed and do not have result columns yet

[eujing]

If an experiment happens to fail on the first trial (maybe quota / provisioning issues) and then MLOS crashes after due to our user scripts expecting a result metric, trying to start the same experiment again gets pandas indexing errors for missing columns at bulk_register.

In such cases, _adjust_signs_df gets called with potentially empty data frames, or ones that do not contain the target opt columns yet because of failed runs.

Then this line will fail with indexing errors:

Maybe the dataframe can be created first before adjusting signs, and then checked if the target columns exist