watch key is CPU100 dead

centos7.2, python2.7 etcd3 0.5.2

I put 20w key to etcd3, try to test python-etcd3 watch performance.

```python
def watch_node():
host = random.choice(ETCD_SERVER)
etcd = etcd3.client(host=host[0], port=host[1])
watch_count = 0
events_iterator, cancel = etcd.watch_prefix('/')
f = open('etcdout', 'w')
for event in events_iterator:
# print(watch_count, event)
f.write(str(watch_count))
f.write("\n")
f.flush()
# f.write(str(event))
watch_count += 1
if watch_count > 1000000:
cancel()
```

```go
package main
import "fmt"
import "time"
import (
"log"
"math/rand"
"os"
"golang.org/x/net/context"
"github.com/coreos/etcd/clientv3"
)

func worker(id int, jobs <-chan int, results chan<- int) {
num := rand.Int31n(10)
time.Sleep(time.Duration(num) * time.Second)

cfg := clientv3.Config{
Endpoints: []string{"http://10.10.3.10:2379", "http://10.1.106.39:2379", "http://10.15.12.253:2379"},
DialTimeout: 5 * time.Second,
}
cli, err := clientv3.New(cfg)
if err != nil {
log.Fatal(err)
}
for j := range jobs {

// ctx, cancel := context.WithTimeout(context.Background(), 3)
resp, err := cli.Put(context.Background(), fmt.Sprintf("/computed/host_vm/%i/%i/state", id, j), "RUNNING")
// defer cancel()
log.Println("xxxxxxxxxxxxxxxxxxxxxx")
if err != nil {
log.Println(err)
cli, err = clientv3.New(cfg)
if err != nil {
log.Println(err)
}
} else {
// print common key info
_ = resp
log.Printf("Set is done. Metadata is %q\n", resp)
}
time.Sleep(10 * time.Second)
results <- j
}
// for j := range jobs {
// fmt.Println("worker", id, "started job", j)
// time.Sleep(time.Second)
// fmt.Println("worker", id, "finished job", j)
// results <- j * 2
// }
}

func main() {
f, err := os.OpenFile("testlogfile", os.O_RDWR | os.O_CREATE | os.O_APPEND, 0666)
if err != nil {
fmt.Println("error opening file: %v", err)
}
defer f.Close()

log.SetOutput(f)

jobs := make(chan int, 500000)
results := make(chan int, 500000)

for w := 1; w <= 30000; w++ {
go worker(w, jobs, results)
}

for j := 1; j <= 500000; j++ {
jobs <- j
}
close(jobs)

for a := 1; a <= 500000; a++ {
<-results
}
}
```

```
(gdb) bt
#0 0x00007f3ab2d4e790 in sem_wait () from /lib64/libpthread.so.0
#1 0x00007f3ab30697b5 in PyThread_acquire_lock () from /lib64/libpython2.7.so.1.0
#2 0x00007f3ab306d442 in lock_PyThread_acquire_lock () from /lib64/libpython2.7.so.1.0
#3 0x00007f3ab303daa4 in PyEval_EvalFrameEx () from /lib64/libpython2.7.so.1.0
#4 0x00007f3ab303f0bd in PyEval_EvalCodeEx () from /lib64/libpython2.7.so.1.0
#5 0x00007f3ab303d76f in PyEval_EvalFrameEx () from /lib64/libpython2.7.so.1.0
#6 0x00007f3ab303f0bd in PyEval_EvalCodeEx () from /lib64/libpython2.7.so.1.0
#7 0x00007f3ab303d76f in PyEval_EvalFrameEx () from /lib64/libpython2.7.so.1.0
#8 0x00007f3ab2fc1908 in gen_send_ex.isra.0 () from /lib64/libpython2.7.so.1.0
#9 0x00007f3ab303b4bd in PyEval_EvalFrameEx () from /lib64/libpython2.7.so.1.0
```

Start with the watch_prefix('/') call in the supplied Python reproducer and compare its behavior with the accompanying Go workload and gdb backtrace. Reproduce the stalled watch under the stated Python, etcd3, and CentOS versions, then identify and address the cause of the CPU hang; done means the watch processes the workload or fails with a clear, reproducible error.

Written by the indexing model from the issue text.