observer_stats: use GetNumPossibleCPUs()

runtime.NumCPU() returns the number of CPUs available to the process.
This excludes offline CPUs, but also does not take into acount task CPU
affinity.

For example:

```
$ cat numcpus.go
package main

import (
        "fmt"
        "runtime"
)

func main() {
        fmt.Printf("%d\n", runtime.NumCPU())
}
$ go run numcpus.go
16
$ taskset --cpu-list 0 go run numcpus.go
1
```

This leads to memory corruption because we pass to the bpf() call a
buffer smaller than the write it will perform, leading to data being
overwritten. Reducing the timeout of the goroutine that reads stat maps
will reproduce this quickly and lead to SIGSEGVs, corrupted prometheus
metrics or other symptoms.

Signed-off-by: Kornilios Kourtis <[email protected]>

pkg/observer/observer_stats.go

@@ -28,7 +28,7 @@ func (k *statKey) String() string            { return fmt.Sprintf("key=%d", k.Ke
 func (k *statKey) GetKeyPtr() unsafe.Pointer { return unsafe.Pointer(k) }
 func (k *statKey) NewValue() bpf.MapValue {
 	return &statValue{
-		Value: make([]int64, runtime.NumCPU()),
+		Value: make([]int64, bpf.GetNumPossibleCPUs()),
 	}
 }
 func (k *statKey) DeepCopyMapKey() bpf.MapKey { return &statKey{k.Key} }
@@ -41,7 +41,7 @@ func (s *statValue) GetValuePtr() unsafe.Pointer {
 }
 func (s *statValue) DeepCopyMapValue() bpf.MapValue {
 	v := &statValue{
-		Value: make([]int64, runtime.NumCPU()),
+		Value: make([]int64, len(s.Value)),
 	}
 	copy(v.Value, s.Value)
 	return v