fix(datascience): glirel_load_model compatible con huggingface_hub 1.x

GLiREL declara proxies/resume_download como required-keyword en
_from_pretrained, pero huggingface_hub 1.x dejo de pasarlos en su
from_pretrained. Aplicamos un classmethod monkey-patch idempotente
que inyecta valores neutros si faltan. Verificado contra glirel==1.2.1
y huggingface_hub==1.13.0 con jackboyla/glirel-large-v0.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

python/functions/datascience/glirel_load_model.py

@@ -56,6 +56,18 @@ def glirel_load_model(
             "`uv pip install glirel` o `uv pip install -e '.[nlp]'`."
         ) from exc
 
+    # huggingface_hub 1.x ya no pasa `proxies`/`resume_download` a
+    # `_from_pretrained`, pero glirel todavia los declara required-keyword.
+    # Wrappeamos el classmethod para inyectar valores neutros si faltan.
+    if not getattr(GLiREL, "_fn_registry_kwargs_patched", False):
+        _orig = GLiREL._from_pretrained.__func__
+        def _patched(cls, **kw):
+            kw.setdefault("proxies", None)
+            kw.setdefault("resume_download", False)
+            return _orig(cls, **kw)
+        GLiREL._from_pretrained = classmethod(_patched)
+        GLiREL._fn_registry_kwargs_patched = True
+
     model = GLiREL.from_pretrained(model_name)
     if hasattr(model, "to"):
         model.to(resolved_device)