Tidy the NIO output-option mapper and finish translating open errnos

[yogthos]

Follow-ups to #897, in two parts.

1. Finish the errno translation (the actual fix)

#897 gave nio-open-output-port a guard with arms for EEXIST and ENOENT and an
else that re-raised. Everything else escaped as the raw Chez condition, which
jolt rendered as a bare java.io.IOException:

(Files/newOutputStream <a directory>)
=> java.io.IOException
   "open-file-output-port: failed for /tmp/d: is a directory"

The JVM answers java.nio.file.FileSystemException reading /tmp/d: Is a directory. That is what #897's class-hierarchy entry for FileSystemException
was for — nothing had reached it yet.

The contract is UnixException.translateToIOException: ENOENT, EEXIST and EACCES
each get a class and carry only the path as their message, and every remaining
errno (EISDIR, ELOOP, ENOTDIR, ENOSPC) is a plain FileSystemException whose
message is <path>: <reason>. Chez hands us the strerror text as the second
irritant of the &i/o-filename condition, so the rendering is byte-identical on
Linux and macOS — both spell EISDIR Is a directory and ENOTDIR Not a directory. The reason is picked by scanning the irritants for a string that is
not the filename rather than by position, so an unexpected irritant list degrades
to the bare path instead of mislabeling the error.

This is strictly a refinement for callers: FileSystemException is an
IOException in the hierarchy, so a catch java.io.IOException that caught the
old escape still catches these.

The new row covers EISDIR and ENOTDIR through both newOutputStream and write,
asserting the class, the IOException relationship, and the message's shape —
<path>: <non-empty reason>. It stops short of pinning the reason text: that is
libc's strerror, not jolt's to promise, and the unit gate runs only on Linux (the
macOS job is the flake workflow's packaging smoke and never invokes make unit),
so an exact-wording row could only rot unseen on macOS.

It also leaves the EACCES arm uncovered: provoking it needs a mode-500 directory,
which does not fail for a suite run as root, and a gate row whose outcome depends
on the runner's uid is worse than an uncovered one-line mapping.

2. Cleanups and a test gap

• nio-write! is dead. Its only caller was Files/write, which fix(nio): honor output stream open options #897 moved
  onto the shared option mapper. nio-write-bv! (still used by copy) and
  nio-output-data->bv cover what it did.
• The Files/write lambda shadowed Chez's file-options macro with a let*
  variable of the same name. It reads correctly today because nothing below the
  binding wants the macro — but a later (file-options no-fail) added inside that
  body would silently become an application of a fixnum. Renamed to fopts.
• APPEND + TRUNCATE_EXISTING now reports the JDK's own wording,
  APPEND + TRUNCATE_EXISTING not allowed.
• CREATE_NEW's refusal was half-covered. It was asserted for plain
  CREATE_NEW and for a symlink, but the other legal combinations —
  CREATE_NEW+APPEND, CREATE_NEW+TRUNCATE_EXISTING, CREATE_NEW+CREATE — were
  exercised only against a missing path, i.e. the create side, so the
  direction the fix is about went unchecked for them. Those take a different
  file-options branch than plain CREATE_NEW does ((no-truncate append) vs
  (file-options)), and what makes them refuse is that neither branch carries
  no-fail, so the open is O_EXCL. The new row pins that each must throw
  FileAlreadyExistsException and leave the existing bytes intact.

Validation

• make unit: 1688/1688 (1686 before this branch, plus two new rows)
• make smoke: 200 passed, 0 failed