main: refuse to write the output into the input file

[woahwhattheheck]

Disclosure, per this repository's AGENTS.md: I am an LLM (Claude), submitting on behalf of the account owner. I am available to discuss this change and to revise it in response to review feedback.

Fixes #427.

scrypt enc secret.txt secret.txt opens secret.txt for reading, reads a
passphrase, then opens the same path with fopen(..., "wb") — which truncates
it. Nothing has been read from the input yet, so the first fread() in
scryptenc_file() returns 0, ferror() is false, and the empty read is
indistinguishable from a legitimately empty file. scryptenc_file() writes the
header and HMAC, returns SCRYPT_OK, and scrypt exits 0. The user's file is
replaced by a 128-byte encryption of nothing, silently.

scrypt dec f f loses the ciphertext the same way, and both are reachable with
the input on stdin (scrypt enc - f < f).

The change

same_file() compares the already-open input against the output path with
fstat()/stat() and reports whether they are the same regular file:

• It uses fstat(fileno(infile)) rather than stat(infilename) so that the
  stdin cases are covered too, and so that symlinks, hard links and different
  spellings of the same path are all handled.
• Only regular files count — opening a character device such as /dev/null
  for writing discards nothing, so scrypt enc /dev/null /dev/null keeps
  working.
• If either stat fails the answer is "different", so the normal case where
  the output file does not exist yet costs one failed stat() and nothing
  else.

The check sits immediately after the input is opened and before the
passphrase is read, so the user is not asked to type a passphrase for a command
that cannot succeed. It jumps to the existing err1 label, which closes the
input; no output file is created.

fileno() is already used in libcperciva/util/readpass.c and <sys/types.h>
in lib-platform/util/memlimit.c, so this adds no new portability burden.

Also

• tests/12-same-file.sh covers enc and dec with the same file, asserts
  exit status 1, the error message, and that the input is byte-for-byte
  unchanged; plus a check that encrypting to a different file still works, so
  a false positive in same_file() would be caught.
• scrypt.1 documents the restriction.

Note on the test number

The test is numbered 12 rather than 11 so that it does not collide with
tests/11-info.sh from #426. The ordering works whichever of the two lands
first; the only overlap is the adjacent line each adds to EXTRA_DIST in
Makefile.am. Happy to renumber if you would rather not take #426.