Record the grid_volume in structure.h5 and test dump/load consistency

[oskooi]

Related to #3282.

Context

dump_structure and load_structure use the same grid_volume at any resolution — but only by construction, and nothing was enforcing it.

Neither Python wrapper builds a gv. Both call through to self.structure (python/simulation.py:2345, :2361), and there is exactly one gv construction path — _init_structure → _create_grid_volume (:2053, :1681) — driven by (cell_size, resolution, dimensions, geometry_center, k_point). The (int)(size*a + 0.5) rounding in volone/vol3d/volcyl (src/vec.cpp:904-941) is deterministic, so identical inputs give a bit-identical gv on both sides. The delayed-load path also runs in the right order, after load_chunk_layout.

The gap was that structure::dump wrote no geometric metadata, so structure::load could only validate byte counts. Confirmed empirically that dumping a (5, 4) cell and loading it into a (4, 5) cell at the same resolution succeeded silently with transposed data — the per-chunk ntot (251×201 vs 201×251) is identical, so none of the existing checks fired.

Changes

• src/vec.cpp:933 — volcyl tested isinteger(rsize) instead of isinteger(rsize * a). Verified before/after: r=2.5, res=2 now gives 5 pixels with no warning (previously warned), and r=3, res=2.5 gives 8 pixels with the warning (previously silent).
• src/structure_dump.cpp — structure::dump now writes a gv_metadata dataset (dim, a, pixel counts, origin; full double precision), and structure::load compares it against the live gv before reading anything, aborting with a message that names the offending field. Files without the dataset load exactly as before.
• python/tests/test_dump_load.py — test_dump_load_grid_volume_matches asserts full gv signature equality across a round trip for 1D/2D/3D/cylindrical at resolutions [10, 13, 17.5, 21.3, 33] (all landing on non-integer pixel counts); plus resolution-mismatch, transposed cell, and legacy-file tests.
• python/tests/test_simulation.py — test_volcyl_pixel_rounding_warning, capturing stderr at the file-descriptor level. The warning tested is from C++ (master_printf_stderr → fprintf(stderr, ...) in src/mympi.cpp), not from Python.
• python/simulation.py + the corresponding block in doc/docs/Python_User_Interface.md — load_structure docstring now states the matching requirements.

Verification

test_dump_load.py: 17 passed single-process and under mpirun -np 2 (which exercises the sharded single_parallel_file=False path). test_simulation.py: 32 passed. C++ tests/dump_load exits 0.

Two Notes:

• @unittest.skipIf(mp.count_processors() > 1, ...) is used rather than the test_dump_load.py's existing mp.with_mpi() guard, because meep::abort throws RuntimeError whenever count_processors() == 1 even in an MPI build (src/mympic.cpp:248). With the existing guard the new tests would never have run here. The pre-existing test's guard is not changed.
• fileds::dump/fields::load (src/fields_dump.cpp) have the identical gap — same byte-count-only checks, no gv record. That is not addressed in this PR.