[executorch][cuda] Find the tuned INT4 kernel's def under its module-level name too

Stack from [ghstack](https://github.com/ezyang/ghstack/tree/0.15.0) (oldest at bottom):
* __->__ #24096

`test_a_tuned_shape_traces_only_its_picked_config` finds the user kernel in Inductor's wrapper by the exact text `def _int4_w4a16_direct_kernel(`. That only matches the wrapper's `async_compile.triton('''...''')` string form, which keeps the kernel's own name.

pytorch/pytorch#200502 makes Inductor's python wrapper define each Triton kernel as module-level code by default. There, the def takes Inductor's kernel name, `_int4_w4a16_direct_kernel_0`. The test failed on its DiffTrain import, D126740082, which was reverted partly for this.

The test now matches `^def <name>\w*\(` and slices back to the `user_autotune(` before that match. The decorator's `configs=[...]` still comes right before the def in both forms, so the check that only the picked config is traced is unchanged. This lands ahead of the #200502 reland.

___

@landed-with-radar-review

Differential Revision: D126992211

Pull Request resolved: #24096

Author: bobrenjc93

backends/cuda/tests/test_int4_quantized_gemm.py

@@ -20,6 +20,7 @@
 import contextlib
 import os
 import pickle
+import re
 import statistics
 import tempfile
 import types
@@ -887,10 +888,10 @@ def linear(x):
             )
         source = "\n".join(code)
         name = autotuner.fn.fn.__name__
-        self.assertIn(f"def {name}(", source)
-        block = source[
-            source.rindex("user_autotune(", 0, source.index(f"def {name}(")) :
-        ]
+        # A module-level kernel's def is renamed to Inductor's kernel name: name plus a suffix.
+        kernel_def = re.search(rf"^def {re.escape(name)}\w*\(", source, re.MULTILINE)
+        self.assertIsNotNone(kernel_def, source)
+        block = source[source.rindex("user_autotune(", 0, kernel_def.start()) :]
         configs = block[: block.index("],")]
         self.assertEqual(configs.count("'num_warps'"), 1)
         _check_close(self, compiled, eager)