Skip 2nd serialization pass of DeviceSerialized

As `"dask"` serialization already converts a CUDA object into headers
and frames that Dask is able to work with, drop code that tries to
serialize frames on host further (as they are already as simple as they
can be). Cuts a fair bit of boilerplate from the spilling path, which
should simplify things a bit.

dask_cuda/device_host_file.py

@@ -13,7 +13,6 @@
 from distributed.utils import nbytes
 from distributed.worker import weight
 
-import numpy
 from zict import Buffer, File, Func
 from zict.common import ZictBase
 
@@ -31,37 +30,35 @@ class DeviceSerialized:
         that are in host memory
     """
 
-    def __init__(self, header, parts):
+    def __init__(self, header, frames):
         self.header = header
-        self.parts = parts
+        self.frames = frames
 
     def __sizeof__(self):
-        return sum(map(nbytes, self.parts))
+        return sum(map(nbytes, self.frames))
 
 
 @dask_serialize.register(DeviceSerialized)
 def device_serialize(obj):
-    headers, frames = serialize(obj.parts)
-    header = {"sub-headers": headers, "main-header": obj.header}
+    header = dict(obj.header)
+    frames = list(obj.frames)
     return header, frames
 
 
 @dask_deserialize.register(DeviceSerialized)
 def device_deserialize(header, frames):
-    parts = deserialize(header["sub-headers"], frames)
-    return DeviceSerialized(header["main-header"], parts)
+    return DeviceSerialized(header, frames)
 
 
 @nvtx_annotate("SPILL_D2H", color="red", domain="dask_cuda")
 def device_to_host(obj: object) -> DeviceSerialized:
     header, frames = serialize(obj, serializers=["dask", "pickle"])
-    frames = [numpy.asarray(f) for f in frames]
     return DeviceSerialized(header, frames)
 
 
 @nvtx_annotate("SPILL_H2D", color="green", domain="dask_cuda")
 def host_to_device(s: DeviceSerialized) -> object:
-    return deserialize(s.header, s.parts)
+    return deserialize(s.header, s.frames)
 
 
 class DeviceHostFile(ZictBase):