make_origs drops columns missing from cleaning_sd once any add_orig fires

[paddymul]

Problem

Both PandasAutocleaning.make_origs (buckaroo/dataflow/autocleaning.py:184) and PolarsAutocleaning.make_origs (buckaroo/polars_buckaroo.py:47) rebuild the cleaned frame solely from cleaning_sd entries. When no column sets add_orig (changed == 0), cleaned_df is returned unchanged. Once any column sets add_orig, the output is constructed only from columns that appear in cleaning_sd with an orig_col_name resolvable in cleaned_df — any column present in cleaned_df but absent from cleaning_sd is silently dropped.

Impact

A cleaning op or command that synthesizes a new column (one with no cleaning_sd entry) loses that column whenever another column triggers add_orig. No error is raised; the column just disappears. The built-in configs don't synthesize columns today, so this is latent, but it constrains what op generators and commands can safely do.

Suggested fix

Iterate cleaned_df.columns as the source of truth and consult cleaning_sd only for the add_orig flag, so columns without an sd entry pass through untouched. Should be folded into the autocleaning op-generation rebuild (#879).

Context

Identified during review of #912; the polars make_origs rewrite there deliberately mirrors the pandas behavior, so both implementations share the hazard.