When freeing memory, use the high-level device getter. (#2029)

`current_device()` can throw an error when the current thread does
not have a context bound, and at the point we free memory there's
no guarantee that a context is bound: The preceding call to
`context()` only primes the *task* local state, and doesn't set
CUDA's *thread* local state. The latter is only set before any
API call that needs it, and `cuCtxGetDevice` is explicitly
exempt from that so that we can check whether a device is bound.

src/pool.jl

@@ -479,7 +479,7 @@ end
       error("Trying to free $buf from an unrelated context")
     end
 
-    dev = current_device()
+    dev = device()
     if stream_ordered(dev)
       # mark the pool as active
       pool_mark(dev)