It's too confusing with a distinct set of IDs for logging.

This adds up to a significant performance gain for gcsfuse, in its sequential
read benchmark.

Also ReadDirOp.

In preparation for allocating them up-front for ReadFile and ReadDir.

The new API lends itself toward reading directly into the output buffer, saving
a memmove and allowing for easy reuse of output buffers.

This eliminates about half of the allocations by bytes for gcsfuse's sequential
read benchmark.