perf: A100 probe results; disable nsys CUDA event trace when supported

- common.nsys_profile_flags(): shared nsys profile flags, adding
  --cuda-event-trace=false on nsys 2025.1+ (on by default there, it widens
  the host gaps between steps; kernel times are unchanged). nsys 2024.5
  rejects the flag, so it is detected from the help text.
- cf/run_probes.py: 3x3, 5x5, 7x7, 9x9 at 1, 4 and 8 streams.
- cf/results/A100/: Merlin7 A100 runs, f32 and f64 with lround, and f32
  with round (104e284) at 20k/2000 and 50k/3000. README lists them.

python/tests/perf/common.py

@@ -37,6 +37,19 @@ def _sh(cmd: str, default: str = "unknown") -> str:
     except Exception:
         return default
 
+def nsys_profile_flags() -> list[str]:
+    """`nsys profile` flags shared by every probe.
+
+    nsys 2025.1 traces CUDA events by default, which widens the host gaps
+    between steps (kernel times are unchanged). Older nsys rejects
+    --cuda-event-trace, so it is added only when the help text lists it.
+    """
+    flags = ["--trace=cuda", "--sample=none", "--cpuctxsw=none",
+             "--force-overwrite=true"]
+    if "--cuda-event-trace" in _sh(f"{NSYS} profile --help 2>&1", default=""):
+        flags.append("--cuda-event-trace=false")
+    return flags
+
 
 def assert_build_fresh(headers: list[str]) -> None:
     """Abort if any header (repo-relative path) is newer than the extension."""