align data passed to write() on 32k boundaries

In donut, OTA installation often encounters the write() system call
doing short writes -- which is legal but unexpected -- or failing with
ENOSPC when plenty of space is available.  Passing aligned memory
buffers to write() appears to prevent (or at least reduce the
frequency) of these problems.  b/1833052 has been filed to look at the
underlying problem, but this change aligns buffers we use with write()
so we can OTA for now (or see if this problem still occurs).