Address PR review comments for layout mapping

Replace hardcoded ROCMLIR_TO_MIOPEN_LAYOUT dict with algorithmic
best-effort conversion: normalize spatial dims (0->H, 1->W), drop
group dim (G), then classify by innermost dimension (channel-first
-> NCHW, channel-last -> NHWC).

- Add "best effort mapping" disclaimer (umangyadav)
- Warn on group convolutions G>1 where dropping G is lossy (dhernandez0)
- Handle K defensively in heuristic (dhernandez0)
- Warn on unrecognised layouts instead of silent passthrough (cursor)
- Run yapf formatting

Made-with: Cursor

Author: dorde-antic

mlir/utils/performance/perfRunner.py

@@ -765,11 +765,9 @@ def benchmark_external(cls, commandline, paths: Paths, arch, num_cu, num_chiplet
         # "Elapsed: " (note the space at the end) and "ms"
         match = ELAPSED_TIME_RE.search(outs)
         if not match:
-            raise RuntimeError(
-                "Failed to parse elapsed time from MIOpenDriver output.\n"
-                "Failing command: " + ' '.join(miopen_driver_cmd) + "\n"
-                "Output:\n" + outs
-            )
+            raise RuntimeError("Failed to parse elapsed time from MIOpenDriver output.\n"
+                               "Failing command: " + ' '.join(miopen_driver_cmd) + "\n"
+                               "Output:\n" + outs)
         elapsed_time_in_ms = match.group(1)
         nanoseconds = float(elapsed_time_in_ms) * 1.0e6
         return config.table_entry(nanoseconds)