Fix race in obtaining index of CU (Xilinx#7522)

Add lock around accessing cu index in core device.

Fixes a potential race when a thread access cu indices while another
thread is updating the indices after an xclbin was loaded.  The race
can only occur when multiple hardware contexts are used because CU
contexts are managed per hardware context by XRT coreutil.

xrt_core::device::get_cuidx lazy gathers cu indices from KMD when a
thread tries to open a CU as part of creating a kernel object.  The
first thread to access the index of a CU can cause all indices to be
updated if the requested CU is not yet registered.

src/runtime_src/core/common/device.cpp

@@ -348,6 +348,10 @@ const std::vector<uint64_t>&
 device::
 get_cus() const
 {
+  // This function returns a reference to internal data
+  // that is modified when an xclbin is loaded.  Normally
+  // not an issue since only single xclbin is supported
+  // when using this API.
   if (m_cu2idx.size() > 1)
     throw error(std::errc::not_supported, "multiple xclbins not supported");
 
@@ -358,6 +362,7 @@ cuidx_type
 device::
 get_cuidx(slot_id slot, const std::string& cuname) const
 {
+  std::lock_guard lk(m_mutex);
   auto slot_itr = m_cu2idx.find(slot);
   if (slot_itr == m_cu2idx.end())
     throw error(EINVAL, "No such compute unit '" + cuname + "'");