Skip coercing to bytes in merge_frames

As the frames we receive are typically mutable, non-`bytes` objects like
`bytearray`s or NumPy `ndarray`s, coercing to `bytes` at this stage
triggers a copy of all frames. As we are going to toss those copied
versions anyways when joining them into a larger `bytes` object, this
ends up being wasteful with memory. Fortunately `bytes.join(...)`
accepts any and all `bytes`-like` objects. So instead just pass them all
through as-is to `bytes.join(...)`, which is free and doesn't require a
copy. Should cutdown on the memory usage in this part of the code.

distributed/protocol/utils.py

@@ -1,7 +1,7 @@
 import struct
 import msgpack
 
-from ..utils import ensure_bytes, nbytes
+from ..utils import nbytes
 
 BIG_BYTES_SHARD_SIZE = 2 ** 26
 
@@ -84,7 +84,7 @@ def merge_frames(header, frames):
         if len(L) == 1:  # no work necessary
             out.extend(L)
         else:
-            out.append(b"".join(map(ensure_bytes, L)))
+            out.append(b"".join(L))
     return out