[ncp] set NCP instance pointer to NULL during pseudo reset (#4065)

This commit fixes the problem with initializing the NCP with
wpantund. The bug was apparent when the NCP image was built with the
FULL_LOGS switch set.

The reason for the bug was that after the pseudo reset the OpenThread
stack wrote logs to NCP's otPlatLog implementation that used the not
yet initialized NCP object. This put the mUartSendTask and
mUpdateChangedPropsTask to the TaskletScheduler linked list.

Later when the NCP itself was initialized, the mNext pointer in the
tasklets mUartSendTask and mUpdateChangedPropsTask was set to null,
causing all tasklets scheduled after the mUartSendTask to be lost. The
NCP reset status could not be in such case sent to the wpantund. The
wpantund then retried resetting the NCP, which repeated the process
until the wpantund gave up.

src/ncp/ncp_base.cpp

@@ -604,7 +604,7 @@ void NcpBase::Log(otLogLevel aLogLevel, otLogRegion aLogRegion, const char *aLog
     otError error  = OT_ERROR_NONE;
     uint8_t header = SPINEL_HEADER_FLAG | SPINEL_HEADER_IID_0;
 
-    VerifyOrExit(!mDisableStreamWrite);
+    VerifyOrExit(!mDisableStreamWrite, error = OT_ERROR_INVALID_STATE);
     VerifyOrExit(!mChangedPropsSet.IsPropertyFiltered(SPINEL_PROP_STREAM_LOG));
 
     // If there is a pending queued response we do not allow any new log
@@ -1172,6 +1172,8 @@ otError NcpBase::CommandHandler_RESET(uint8_t aHeader)
         mUpdateChangedPropsTask.Post();
     }
 
+    sNcpInstance = NULL;
+
     return error;
 }