Use memoryview in unpack_frames (#3980)

* Minor whitespace adjustment

* Coerce input to `memoryview` in `unpack_frames`

Selecting out each frame from the input causes a copy, which increases
memory usage and slows down `unpack_frames`. To fix this, coerce the
input to a `memoryview`. This way slices into the `memoryview` only take
a view onto the underlying data, which is quite fast and doesn't result
in additional memory usage.

distributed/protocol/utils.py

@@ -120,8 +120,11 @@ def unpack_frames(b):
     --------
     pack_frames
     """
+    b = memoryview(b)
+
     fmt = "Q"
     fmt_size = struct.calcsize(fmt)
+
     (n_frames,) = struct.unpack_from(fmt, b)
     lengths = struct.unpack_from(f"{n_frames}{fmt}", b, fmt_size)