Fix leaked identifiers in H5DSattach_scale()

[hyoklee]

Describe your changes

H5DSattach_scale() rewrites the scale's REFERENCE_LIST attribute whenever a
dataset is attached to a scale that already has one, and that rewrite acquired
three things it never released:

tmp_id = H5Ropen_object(&ndsbuf[j].ref, H5P_DEFAULT, H5P_DEFAULT);
if (tmp_id < 0)
    goto out;
if (H5Rcreate_object(tmp_id, ".", H5P_DEFAULT, &ndsbuf_w[j].ref) < 0) {
    H5Dclose(tmp_id);      /* error path closed it */
    goto out;
}
                           /* success path did not */

1. the reopened dataset identifier above -- one per reference already attached
   to the scale;
2. the references read from the old attribute into ndsbuf, which were
   free()d without ever being destroyed;
3. the reference appended at index nelmts - 1 of ndsbuf_w, because that
   buffer was reclaimed with sid, the old dataspace, which is one element
   shorter than the buffer.

Each of these keeps the file open, which is the symptom in the issue: a later
H5Fcreate() with H5F_ACC_TRUNC on the same file fails with "unable to
truncate a file which is already open", an error with nothing in it to point
back at a dimension scale. H5DSdetach_scale() closes the identifier and
reclaims both buffers; attach now does the same.

Reachability

The rewrite loop runs only on the new-style reference path, which
H5DSwith_new_ref() selects when the library is built with
H5_DIMENSION_SCALES_WITH_NEW_REF or when the object's terminal VOL
connector is not the native one. Stacking a pass-through connector over the
native connector does not reach it -- H5VLobject_is_native() compares the
terminal connector class, so pass-through-over-native still counts as native.
An earlier revision of this PR said "only a pass-through VOL connector"; the
comments and the H5DSwith_new_ref() documentation are corrected accordingly.

Test

test_attach_scale_id_leak() in hl/test/test_ds.c attaches one scale to four
datasets, then checks that the scale is the only dataset identifier still open
and that the file can be truncated after being closed. It skips itself on the
old-reference path, where the loop opens nothing. Reverting either fix fails
it:

reverted	result
H5Dclose(tmp_id)	7 dataset ID(s) open after attaching, expected 1
the H5Treclaim() calls	the truncate fails
nothing	passes

All 133 tests pass in a -DHDF5_DIMENSION_SCALES_NEW_REF=ON build; HL_* and
the h5dump/h5ls/h5repack tool tests pass in a default build.

CMake

-DHDF5_DIMENSION_SCALES_NEW_REF=ON was a no-op, so the code path above could
not be exercised from a stock build at all. The option was declared in
hl/CMakeLists.txt, a child scope that the top-level H5pubconf.h
configure_file() never sees, so H5_DIMENSION_SCALES_WITH_NEW_REF was never
defined and libhdf5.settings reported nothing for it. It now lives in
CMakeBuildOptions.cmake, which is included before both src/ and hl/.

Issue ticket number (GitHub or JIRA)

Fixes #6639

Checklist before requesting a review

• My code conforms to the guidelines in CONTRIBUTING.md
• I made an entry in release_docs/CHANGELOG.md (bug fixes, new features)
• I added a test (bug fixes, new features)

🤖 Generated with Claude Code

[github-actions]

Review Checklist

This PR touches the following areas. Each needs a sign-off
from its listed owners before merging.

• hl
  • @mattjala (manually added) — approval required
• *.cmake — @lrknox
  • @mattjala (manually added) — approval required
• release_docs — @lrknox

[jhendersonHDF]

This part and below should be moved back to hl/ so that only the option () and mark_as_advanced () are in this file. The changes in #6617 fix the configuration file so that it uses the option value and doesn't need the variables below. In fact, I think setting DIMENSION_SCALES_WITH_NEW_REF can just be removed. I just need to update the usage in H5build_settings.cmake.c.in.