refactor: replace uniform_ noise with rand-based formulation for torch.compile

- Replace `empty_like(...).uniform_()` with `rand_like()`-based expression
- Avoid in-place random ops that are problematic for torch.compile / dynamo
- Preserve identical noise range [-half, half] and statistical behavior

Author: studyingeugene

compressai/entropy_models/entropy_models.py

@@ -158,7 +158,7 @@ def quantize(
 
         if mode == "noise":
             half = float(0.5)
-            noise = torch.empty_like(inputs).uniform_(-half, half)
+            noise = 2*half*torch.rand_like(inputs) - half
             inputs = inputs + noise
             return inputs