[BugFix]: Sparse2of4 example sparsity-only case (#1282)

The Sparse2of4 example supports two modes of operation:

- Sparsity-only
- Sparsity + Quantization (only supported for symmetric quantization)

In a recent update, we removed the `ConstantPruningModifier` from the
Sparse2of4 example and added a check to raise an error if **asymmetric
quantization** was being used. However, this check was incorrectly
placed outside the quantization-specific code path and unconditionally
accessed the scheme attribute from the quantization modifier.

This caused failures in sparsity-only cases, as reported in
[INFERENG-483](https://issues.redhat.com/browse/INFERENG-483).

### Fix
This PR moves the asymmetric quantization check inside the quantization
code path to ensure it is only evaluated when quantization is actually
enabled.

### Testing
The example was tested by running it both with and without the `--fp8`
flag. In both cases, successful completion was verified.

Signed-off-by: Rahul Tuli <rahul@neuralmagic.com>

Author: rahul-tuli

examples/sparse_2of4_quantization_fp8/llama3_8b_2of4.py

@@ -60,13 +60,13 @@ def get_recipe(fp8_enabled):
         )
         save_dir = MODEL_ID.split("/")[1] + "2of4-W8A8-FP8-Dynamic-Per-Token"
 
-    # check that asymmetric quantization is not being used
-    q_scheme = base_recipe[1].scheme
-    if not isinstance(q_scheme, str) and not q_scheme["weights"].symmetric:
-        raise ValueError(
-            "Asymmetric quantization with 2of4 sparsity is not supported by vLLM. "
-            "Please use symmetric quantization"
-        )
+        # check that asymmetric quantization is not being used
+        q_scheme = base_recipe[1].scheme
+        if not isinstance(q_scheme, str) and not q_scheme["weights"].symmetric:
+            raise ValueError(
+                "Asymmetric quantization with 2of4 sparsity is not supported by vLLM. "
+                "Please use symmetric quantization"
+            )
 
     return base_recipe, save_dir