Improve benchmarks/dynamo:check_perf_csv output and failure summary (#161728)

Resolves https://github.com/pytorch/pytorch/issues/161290

## Summary

Expands `dynamo/check_perf_csv.py` output capabilities with latency, compile time and memory information:

- Display's measured speedup and display % from target
- Added clear messaging for all passing model tests when no regression is found
- Added error handling if csv file is missing

### Example (Failing Check)

```bash
python benchmarks/dynamo/check_perf_csv.py -f reports-dir/inductor_training_smoketest.csv -t 1.40
```

**Example Output:**
```
Checking inductor_training_smoketest.csv (speedup threshold >= 1.40x)
hf_Bert                            speedup=1.005x, latency=390.8 ms/iter, compile=1.526s, mem_ratio=1.02x (eager=360.6 GB, dynamo=369.3 GB)
Error 1 model(s) performance regressed
    hf_Bert
  - hf_Bert: 1.005x (< 1.40x; -28.2% from target)
```

### Example (Passing Check)

```bash
python benchmarks/dynamo/check_perf_csv.py -f reports-dir/inductor_training_smoketest.csv -t 1.40
```

**Example Output:**
```
Checking inductor_training_smoketest.csv (speedup threshold >= 1.00x)
hf_Bert                            speedup=1.005x, latency=390.8 ms/iter, compile=1.526s, mem_ratio=1.02x (eager=360.6 GB, dynamo=369.3 GB)
All 1 model(s) passed threshold check (>= 1.00x)
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/161728
Approved by: https://github.com/isuruf

benchmarks/dynamo/check_perf_csv.py

@@ -9,28 +9,61 @@ def check_perf_csv(filename, threshold, threshold_scale):
     """
     Basic performance checking.
     """
+    try:
+        df = pd.read_csv(filename)
+    except FileNotFoundError:
+        print(f"Error: File {filename} not found")
+        sys.exit(1)
 
-    df = pd.read_csv(filename)
+    effective_threshold = threshold * threshold_scale
+    print(f"Checking {filename} (speedup threshold >= {effective_threshold:.2f}x)\n")
 
     failed = []
     for _, row in df.iterrows():
         model_name = row["name"]
-        speedup = row["speedup"]
-        if speedup < threshold * threshold_scale:
-            failed.append(model_name)
+        speedup = float(row["speedup"])
+        abs_latency = float(row["abs_latency"])
+        compilation_latency = float(row["compilation_latency"])
+        compression_ratio = float(row["compression_ratio"])
+        eager_peak_mem = float(row["eager_peak_mem"])
+        dynamo_peak_mem = float(row["dynamo_peak_mem"])
 
-        print(f"{model_name:34} {speedup}")
+        perf_summary = f"{model_name:34} speedup={speedup:.3f}x"
+        if pd.notna(abs_latency):
+            perf_summary += f", latency={abs_latency:.1f} ms/iter"
+        if pd.notna(compilation_latency):
+            perf_summary += f", compile={compilation_latency:.3f}s"
+        if pd.notna(compression_ratio):
+            perf_summary += f", mem_ratio={1 / compression_ratio:.2f}x"
+            if pd.notna(eager_peak_mem) and pd.notna(dynamo_peak_mem):
+                perf_summary += (
+                    f" (eager={eager_peak_mem:.1f} GB, dynamo={dynamo_peak_mem:.1f} GB)"
+                )
+
+        if speedup < effective_threshold:
+            failed.append((model_name, speedup))
+
+        print(perf_summary)
 
     if failed:
         print(
             textwrap.dedent(
                 f"""
-                Error {len(failed)} models performance regressed
-                    {" ".join(failed)}
+                Error {len(failed)} model(s) performance regressed
+                    {" ".join([name for name, _ in failed])}
                 """
             )
         )
+        for name, sp in sorted(failed, key=lambda x: x[1]):
+            pct_from_target = (sp / effective_threshold - 1.0) * 100.0
+            print(
+                f"  - {name}: {sp:.3f}x (< {effective_threshold:.2f}x; {pct_from_target:.1f}% from target)"
+            )
         sys.exit(1)
+    else:
+        print(
+            f"\nAll {len(df)} model(s) passed threshold check (>= {effective_threshold:.2f}x)"
+        )
 
 
 if __name__ == "__main__":
@@ -44,7 +77,7 @@ if __name__ == "__main__":
         "-s",
         type=float,
         default=1.0,
-        help="multiple threshold by this value to relax the check",
+        help="multiply threshold by this value to relax the check",
     )
     args = parser.parse_args()
     check_perf_csv(args.file, args.threshold, args.threshold_scale)