increase xperf and tracelog buffer sizes

Otherwise a lot of the benchmarks trigger errors and warnings about lost
events, for example (pardon my french ...xperf):

```
xperf: error: rustc-perf-counters: Le nom d'instance passé n'a pas été
reconnu valide par un fournisseur de données WMI. (0x1069).
xperf: warning: Session "rustc-perf-counters" lost 5204 events.
```

Author: lqd

collector/src/rustc-fake.rs

@@ -129,6 +129,9 @@ fn main() {
                 // falling back to assuming it's on the PATH.
                 let xperf = std::env::var("XPERF").unwrap_or("xperf.exe".to_string());
                 let mut cmd = Command::new(&xperf);
+                cmd.args(&["-MinBuffers", "512"])
+                    .args(&["-MaxBuffers", "3000"])
+                    .args(&["-BufferSize", "1024"]);
                 assert!(cmd.output().is_ok(), "xperf.exe could not be started");
 
                 // Go ahead and run `xperf -stop rustc-perf-counters` in case there are leftover
@@ -150,6 +153,12 @@ fn main() {
                     "CSWITCH+PROC_THREAD+LOADER",
                     "-PMC",
                     "InstructionRetired,TotalCycles:CSWITCH",
+                    "-b",
+                    "1024",
+                    "-min",
+                    "512",
+                    "-max",
+                    "2048",
                 ]);
                 let status = cmd.status().expect("failed to spawn tracelog");
                 assert!(status.success(), "tracelog did not complete successfully");