Add atol to test_within_transformer output comparison (#24723)

Summary:
`test_within_transformer` compares the static-attention transformer against the MHA transformer with `torch.isclose(rtol=1e-3)` and the default `atol=1e-8`, which is effectively a pure relative tolerance. Some output logits land near zero (~7e-4), where float noise of ~3e-7 is already a large fraction of the relative budget. The margin depends on which GEMM kernel MKL picks: on master the worst relative error is 3.2e-4 to 6.3e-4 with the default path and 9.5e-4 with `MKL_CBWR=COMPATIBLE`, just under the 1e-3 limit.

D129776559 (pytorch/pytorch#197785) makes `Vectorized<float>::exp_u20()`, used by the CPU SDPA softmax, more accurate, which shifts the MHA output enough to cross the limit on CI hosts.

Add `atol=1e-5`, matching `test_sliding_window_chunked_prefill` in the same file. The largest absolute error observed is 3.6e-7, while a real attention regression moves outputs far more than 1e-5. This is independent of the PyTorch change and can land before the DiffTrain import.

Authored with assistance from Claude.


Reviewed By: georgehong

Differential Revision: D130128860

Pulled By: huydhn

Author: huydhn

examples/models/llama/tests/test_static_attention.py

@@ -250,7 +250,9 @@ def test(style, attention_type):
                 ys.append(y_i)
 
             self.assertTrue(
-                torch.isclose(ys[-1].flatten(), expected.flatten(), rtol=1e-3).all()
+                torch.isclose(
+                    ys[-1].flatten(), expected.flatten(), rtol=1e-3, atol=1e-5
+                ).all()
             )
 
         for args in itertools.product(