Change olmOCR2 temperature from 0.1 to 0.0 to match original implementation

- Update default temperature to 0.0 (was 0.1)
- Matches olmOCR pipeline.py default (build_page_query uses 0.0)
- Should help reduce repetitive generation issues
- Add olmOCR2 documentation to README

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

olmocr2-vllm.py

@@ -263,7 +263,7 @@ def main(
     model: str = "allenai/olmOCR-2-7B-1025-FP8",
     max_model_len: int = 16384,
     max_tokens: int = 8192,
-    temperature: float = 0.1,
+    temperature: float = 0.0,
     gpu_memory_utilization: float = 0.8,
     hf_token: str = None,
     split: str = "train",
@@ -283,7 +283,7 @@ def main(
         model: HuggingFace model ID for olmOCR
         max_model_len: Maximum context length
         max_tokens: Maximum tokens to generate per image
-        temperature: Sampling temperature (0.1 for deterministic)
+        temperature: Sampling temperature (0.0 for deterministic)
         gpu_memory_utilization: Fraction of GPU memory to use
         hf_token: HuggingFace token for authentication
         split: Dataset split to process
@@ -334,7 +334,7 @@ def main(
         limit_mm_per_prompt={"image": 1},
     )
 
-    # Sampling parameters - olmOCR uses low temperature for deterministic output
+    # Sampling parameters - olmOCR uses temperature 0.0 for deterministic output
     sampling_params = SamplingParams(
         temperature=temperature,
         max_tokens=max_tokens,
@@ -508,8 +508,8 @@ Examples:
     parser.add_argument(
         "--temperature",
         type=float,
-        default=0.1,
-        help="Sampling temperature (default: 0.1 for deterministic output)",
+        default=0.0,
+        help="Sampling temperature (default: 0.0 for deterministic output)",
     )
     parser.add_argument(
         "--gpu-memory-utilization",