Extra fix needed for @vicatcu's PR of @chegewara's work See below

So I merged the pull-request mentioned in espressif#4046, and noticed a regression
compared to the code currently in master:

In my app I use light-sleep, and to do so I have to carefully not
just turnoff bluetooth, but stop and delete my various services before
going to sleep. During this process I would call BLEDevice::deinit(false).
That still works fine.

However, when my app woke from light-sleep and needed to restart bluetooth
it was hanging in registerApp:

```
Starting bluetooth
[D][BLEDevice.cpp:80] createServer(): >> createServer
[D][BLEServer.cpp:291] registerApp(): >> registerApp - 1
[D][FreeRTOS.cpp:164] take(): Semaphore taking: name: RegisterAppEvt (0x3ffdd7b0), owner: <N/A> for registerApp
[D][FreeRTOS.cpp:173] take(): Semaphore taken:  name: RegisterAppEvt (0x3ffdd7b0), owner: registerApp
```

To I poked around a bit, and it seems that in the PR espressif#4050 branch, there
is a change compared to master.  A bit of code that used to be #ifdef
was changed to be #ifndef.  Which has the effect of never setting
initialized to false.  If initialized is left true, my call to BLEDevice::init()
was having no effect.

So I've changed this ifdef back to what master had, and now my app works
nicely and the corruption I was previously seeing in my BLE structures
seems fixed (the original goal of pulling in 4050.

Author: geeksville

Diff for: libraries/BLE/src/BLEDevice.cpp

@@ -620,7 +620,7 @@ void BLEDevice::removePeerDevice(uint16_t conn_id, bool _client) {
     esp_bluedroid_deinit();
     esp_bt_controller_disable();
     esp_bt_controller_deinit();
-#ifndef ARDUINO_ARCH_ESP32
+#ifdef ARDUINO_ARCH_ESP32
     if (release_memory) {
         esp_bt_controller_mem_release(ESP_BT_MODE_BTDM);  // <-- require tests because we released classic BT memory and this can cause crash (most likely not, esp-idf takes care of it)
     } else {