[QAT Lora 6/N] torch.compile for reference implementation (#3343)

### Changes

- Added `torch.compile` for `forward` & `backward` in reference
implementation.

### Reason for changes

- Training speed-up from 7 minutes to 5 minutes for 1 epoch of phi3.5
qat-lora tuning

### Related tickets

- 163973

### Tests

- TBD

Author: nikita-malininn

nncf/torch/quantization/reference.py

@@ -121,5 +121,5 @@ def tune_range(
 
 class ReferenceQuantizedFunctions:
     _executor = ReferenceQuantize(backend_type=ReferenceBackendType.TORCH)
-    Quantize_forward = _executor.forward
-    Quantize_backward = _executor.backward
+    Quantize_forward = torch.compile(_executor.forward)
+    Quantize_backward = torch.compile(_executor.backward)