This allows to remove platform-specific code and avoids it being
maintained in this codebase.

This is a refactor, apart from potentially using one more file
descriptor at runtime, due to the same file being opened already
earlier. However, the increase in file descriptors will be undone in the
next commit.

* This removes the requirement from the call site to use a raw C-style
  FILE pointer. Now the call site can use any style of file IO.
* This makes the edge-case impossible to call FileCommit with nullptr,
  which would eventually lead to a segfault.
* This removes the burden from the call site to first open the file.
* This avoids a fseek if the file was opened using FlatFileSeq::Open
  with a non-zero nPos.
* Where the file was already open at a call site, it removes the burden
  to keep it open for FileCommit. Call sites are now expected to close
  or flush the file before calling FileCommit.
* This removes the requirement from the call site to open the file with
  the right mode (write).

This should be a refactor, because the following corner cases never happen:
* FlatFileSeq::Open has a side-effect to create the file when it is
  missing. Thus, in WriteFilterToDisk, if the first filter overflew the
  first file, the node would abort. Previously it would create an empty
  first filter file.