fix: add PyTorch memory allocator config to prevent fragmentation

Root Cause Analysis:
- 256h context reduction DID work: PyTorch allocation dropped from 20.42 GB to 7.98 GB
- OOM persisted due to memory fragmentation: 12.61 GB reserved but unallocated
- Error: Could not allocate 12.44 GB contiguous block for attention matrix
- PyTorch error message: "try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True"

Fix Applied:
- Set os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True' BEFORE importing torch
- This enables PyTorch's expandable memory segments feature
- Prevents fragmentation by allowing PyTorch to use non-contiguous memory blocks

Expected Outcome:
- The 12.61 GB of reserved memory can now be used for the 12.44 GB attention allocation
- Should fit comfortably within L4 24GB VRAM: 7.98 GB + 12.44 GB = 20.42 GB

Testing Plan:
1. Deploy to HF Space
2. Run smoke test (AT_CZ border, 7-day forecast)
3. If successful: Run full Oct 1-7 forecast (7 borders)
4. Calculate D+1 MAE and verify <150 MW threshold

References:
- PyTorch memory management: https://pytorch.org/docs/stable/notes/cuda.html#environment-variables
- Chronos-2 paper: Context window directly affects attention memory (O(n²))

Co-Authored-By: Claude <[email protected]>

src/forecasting/chronos_inference.py

@@ -10,6 +10,12 @@ import os
 import time
 from typing import List, Dict, Optional
 from datetime import datetime, timedelta
+
+# CRITICAL: Set PyTorch memory allocator config BEFORE importing torch
+# This prevents memory fragmentation issues that cause OOM even with sufficient free memory
+# See: https://pytorch.org/docs/stable/notes/cuda.html#environment-variables
+os.environ['PYTORCH_CUDA_ALLOC_CONF'] = 'expandable_segments:True'
+
 import polars as pl
 import pandas as pd
 import numpy as np