Fix/test minimal lowest direct

[martin-schlipf]

CI: make test-minimal actually test the lowest versions

Summary

test-minimal passed --resolution lowest-direct to uv lock only, but the
following uv sync (and uv run) re-resolve the lockfile as well, so the pins
were discarded before the tests ever ran:

after lock: ipython 8.26.0
after sync: ipython 9.17.1

The job was therefore testing the newest versions, and the declared lower
bounds were never guarded — which is why an IPython 9.17 regression showed up
there at all (#317). Setting UV_RESOLUTION for the whole job makes every uv
invocation agree.

With the job actually doing its stated job, four bounds turn out to be too low
next to numpy>=2.0. Each is raised to the first version that works, and I
checked that one notch below still fails:

dep	was	now	why
h5py	>=3.9.0	>=3.11.0	3.9/3.10 are built against numpy 1.x → numpy.dtype size changed on import
pandas	>=2.0	>=2.2.2	same up to 2.1; 2.2.0 and 2.2.1 declare numpy<2
scipy	>=1.12.0	>=1.15.0	interpolate.AAA, used by _third_party.numeric, was added in 1.15
spglib	>=2.1	>=2.5	get_symmetry_dataset returns a plain dict before 2.5, while _calculation.structure and _calculation.symmetry access it by attribute

core/pyproject.toml carried the same h5py>=3.9.0 next to numpy>=2.0, so it
is raised too. There is no test-minimal job for core, so nothing in CI would
have caught it.

Both lock files are updated because they record the specifiers in
requires-dist; no resolved version changes. I edited just those lines rather
than running uv lock, which with my older local uv would have re-added 41
lines of dependency markers that #316 removed.

Testing

• Fixed job end-to-end: numpy 2.0.0, h5py 3.11.0, pandas 2.2.2, scipy 1.15.0,
  spglib 2.5.0, ipython 8.26.0 → 3248 passed, 90 skipped, 2 xfailed
• test-core scenario (cp core/* . + uv sync --locked --all-extras --dev)
  → 2926 passed, 310 skipped
• Root uv sync --locked succeeds, so test-full / install see no stale lock