create: do not archive an atime we caused ourselves, fixes #6194

[ThomasWaldmann]

borg has to open() a fs item before it can fstat() it. On platforms without O_NOATIME
support (or when we are not allowed to use it), that open() may already have updated the
atime - and then we archive an atime that we caused ourselves, instead of the one the item
had before borg touched it.

stat_update_check() now returns a os.stat_result look-alike (StatOrigAtime) that has
everything from the fd-based stat, but the atime from the stat we did before opening the
item - as suggested in the issue. This is only done if the atime actually changed between
the two stat calls, so when O_NOATIME worked, the plain stat result is returned as before
(no proxy object, no overhead).

Doing this in stat_update_check() covers all its call sites at once: regular files, fifos,
devices and both directory opens.

Note: this only changes what we archive - without O_NOATIME we still can not avoid
updating the atime in the source filesystem.

Tests:

• test_atime_open_updates_atime simulates a platform where the open() already updates the
  atime (by reading a byte in a monkeypatched os_open) and checks that the archived atime is
  still the one from before. It fails without the fix.
• unit tests for stat_update_check(): atime fixup, unchanged atime (returns the stat result
  as is) and the two race condition checks, which had no test coverage.

2nd commit: test_atime: do not create a hardlink

create_test_files creates input/hardlink as a hard link to input/file1, so both share one
inode and anything touching one of them also changes the atime of the other. That is not what
test_atime is about and it made the test fail on cygwin (part of #7218):

• At extraction time, borg creates the hard link via os.link(). On cygwin, os.link() updates
  the atime of the file it links to, at least if that file has the x bit set - and
  create_test_files does chmod(input/file1, 0o4755). Measured: mode 0o644 keeps the atime,
  0o755 and 0o4755 lose it. So the atime borg had correctly restored for input/file1 was
  overwritten by "now" again, and the test saw the wrong value. Tracing the extraction confirms
  that set_times() is called exactly once for input/file1, with the correct atime, and that
  the very next os.link() clobbers it.
• Independently of that: while backing up, borg has to open() input/file1, which updates the
  atime of the shared inode. When borg then gets to input/hardlink, that already changed atime
  is what the stat before the open sees, so stat_update_check() can not recognize it as one we
  caused ourselves and archives it. The test never looks at input/hardlink, so this did not
  make it fail, but hardlinks and atime testing simply do not mix well.

So the test now just does not create the hardlink.

Note for reviewers: on cygwin it is this 2nd commit that makes test_atime pass - the 1st commit
does not change anything there, because cygwin does not show the atime update in the fstat()
right after the open(), so the fixup never triggers. The 1st commit is for the platforms where
it does.

Testing:

• cygwin (win11, py3.12): test_atime failed reliably before, passes now (3 of 3 runs, both
  archiver and remote_archiver). Full extract_cmd_test.py: 59 passed, 56 skipped.
• FreeBSD 15.1 (py3.14.7), rebased onto current master: full test suite 2643 passed, 975 skipped.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87.00%. Comparing base (6cef09f) to head (00fd667).
⚠️ Report is 17 commits behind head on master.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##           master   #10111      +/-   ##
==========================================
+ Coverage   86.94%   87.00%   +0.05%     
==========================================
  Files         101      101              
  Lines       17983    17993      +10     
  Branches     2737     2738       +1     
==========================================
+ Hits        15636    15654      +18     
+ Misses       1639     1635       -4     
+ Partials      708      704       -4     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.