Fill missing categoricals as UNKNOWN instead of the string "nan"

[nikhilsi]

What was wrong

In openavmkit/cleaning.py, _fill_unknown_values (lines 748-749 at
7952236) casts categorical fields to str before filling:

df[field] = df[field].astype("str")
df[field] = df[field].fillna("UNKNOWN")

astype("str") converts NaN to the literal string "nan" first, so the
fillna on the next line never finds anything to fill. Models end up with
a "nan" category instead of the intended "UNKNOWN".

The change

Reorder so the fill happens while the nulls are still real:

df[field] = df[field].astype("object").fillna("UNKNOWN").astype("str")

Going through object keeps NaN intact for the fill and also sidesteps the
categorical-dtype restriction on filling with an unseen category.

What changes downstream

The category label for missing values changes from "nan" to "UNKNOWN", so
one-hot column names and tree categorical levels shift accordingly.
Refitting after upgrading will produce slightly different encodings than
cached models built against the "nan" label; model caches should be
invalidated once this lands.

Verification

A test is included. It runs a frame containing NaN in a declared
categorical field through the fill step and asserts the result contains
"UNKNOWN" and no "nan" values. At 7952236 the "nan" assertion fails; with
this change it passes.

Fixes #377.

[github-actions]

Thank you for your contribution.
Please sign our CLA at the following link:
Click here to sign the CLA.

No action is required from you in this PR thread. Once you have signed the CLA externally, a maintainer will verify your signature and record it here on your behalf by commenting:

---

I affirm that this contributor has signed the CLA

---

_{You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot.}