Add symmetric option to group_structures, pass through to sm.fit() and test

[kavanase]

Summary

This is a small PR to add the symmetric option to StructureMatcher.group_structures(), which is then passed through to the underlying StructureMatcher.fit() usage.

Summary with the help of Claude:
StructureMatcher.fit is directional; symmetric=False by default, and its docstring says so. But group_structures consumes it as if it were an equality test: it pops a reference and always passes that reference as struct1. So which structure happens to come first decides the grouping.

This repo already keeps a structure pair for the asymmetry, used by test_fit:

# src/pymatgen/util/testing.py:36
VASP_IN_DIR: str = f"{TEST_FILES_DIR}/io/vasp/inputs"

sm = StructureMatcher(comparator=ElementComparator(), ltol=0.6, stol=0.6, angle_tol=6)
s1 = Structure.from_file(f"{VASP_IN_DIR}/POSCAR_fit_symm_s1")
s2 = Structure.from_file(f"{VASP_IN_DIR}/POSCAR_fit_symm_s2")

len(sm.group_structures([s1, s2]))                   # 1 group
len(sm.group_structures([s2, s1]))                   # 2 groups   <-- same inputs
len(sm.group_structures([s1, s2], symmetric=True))   # 2 groups
len(sm.group_structures([s2, s1], symmetric=True))   # 2 groups

Using test-files/io/vasp/inputs/POSCAR_fit_symm_s1 and _s2, from tests/core/test_structure_matcher.py:529-537.

This PR threads a symmetric kwarg through to the fit call so callers that need an order-independent
grouping can ask for one. It defaults to False, so no existing caller changes behaviour, and it is
ignored under anonymous=True since fit_anonymous has no symmetric mode. One test added.

Motivation: MaterialsProjectDFTMixingScheme groups GGA and r2SCAN structures this way, and the
asymmetry makes its phase diagrams depend on PYTHONHASHSEED — see materialsproject/pymatgen#3113 and
the follow-up PR (I'll be raising now) to materialsproject/pymatgen#4703, which passes symmetric=True here. Over the 120 permutations of the pre-group in that bug, the current default yields 3 distinct groupings and symmetric=True yields 1, at 1.04x runtime for the full pipeline.

Note this does not make greedy grouping order-independent in general — a symmetric relation can still be non-transitive — but it removes the directional class of it at least.

Checklist

• Tests for the affected code pass locally (e.g. uv run pytest tests/core -k lattice
  for a lattice change — the full suite runs in CI)
• Lint passes: uv run ruff check . && uv run ruff format --check .

Breaking changes?

• This PR contains breaking changes (removals/renames, signature or
  behavior changes, changed defaults/output, dropped Python versions, ...).
  If so, prefix the title with [breaking] and describe the change and
  migration steps under ## Breaking Changes below.

Breaking Changes

[shyuep]

🤖 Automated PR review generated by Claude (posted via @shyuep's scheduled review routine)

Clean, minimal change. The order-dependence of group_structures from directional fit() is a real and long-standing footgun (cf. materialsproject/pymatgen#3113), and threading symmetric through with a False default preserves existing behavior. Docstring correctly notes the kwarg is ignored under anonymous=True and that greedy grouping remains non-transitive in general.

Test reuses the existing asymmetric POSCAR pair and covers both orderings — good. All CI green.

One minor note: symmetric=True roughly doubles the fit cost per pair; the 1.04x pipeline figure in the description is reassuring, so no action needed. LGTM.

[shyuep]

Thanks.