fix: reduce context window to 256h to fit T4 GPU memory

CUDA OOM Error Fix:
- Previous: 512h context × 615 features = 12.44 GiB allocation
- New: 256h context × 615 features = fits within 14.74 GiB T4 GPU
- Still provides 10+ days of historical context for forecasting

Changed:
- chronos_inference.py: context_hours default 512 -> 256
- dynamic_forecast.py: context_hours default 512 -> 256

This enables multivariate covariate forecasting with all 615 features
on free-tier T4 GPU (14.74 GB memory).

Error: OutOfMemoryError when allocating 12.44 GiB with 512h context
Fix: Reduced to 256h, maintaining sufficient historical signal

src/forecasting/chronos_inference.py

@@ -105,7 +105,7 @@ class ChronosInferencePipeline:
         run_date: str,
         borders: Optional[List[str]] = None,
         forecast_days: int = 7,
-        context_hours: int = 512,
+        context_hours: int = 256,
         num_samples: int = 20
     ) -> Dict:
         """

src/forecasting/dynamic_forecast.py

@@ -30,7 +30,7 @@ class DynamicForecast:
     def __init__(
         self,
         dataset: pl.DataFrame,
-        context_hours: int = 512,
+        context_hours: int = 256,
         forecast_hours: int = 336  # Fixed at 14 days
     ):
         """