fix: widen ravel_multi_index to int64 on overflow

[andrinr]

Summary

NumpyBackend.ravel_multi_index unconditionally cast the linearised result to the caller's multi_index.dtype. When that dtype was int32 and the shape product exceeded 2**31, the cast silently wrapped to negative. Downstream callers — primarily CompressedSparseMatrix.compress in phiml/math/_sparse.py — then surface it as:

ValueError: index <negative> is out of bounds for array with size <product>

from np.unravel_index on the next line.

Fix: preserve the caller's dtype when it fits, widen to int64 when np.prod(shape) > np.iinfo(dtype).max. The JAX backend already doesn't downcast, so the patch brings the numpy backend in line.

Reproducer

import numpy as np
from phiml.backend._numpy_backend import NUMPY

indices = np.array([[40000, 40000]], dtype=np.int32)
shape = (200_000, 200_000)   # product = 4e10 > 2**31

lin = NUMPY.ravel_multi_index(indices, shape)
# before this PR:
#   lin.dtype == int32, lin == [-589894592]
#   NUMPY.unravel_index(lin, shape)
#     → ValueError: index -589894592 is out of bounds for array with size 40000000000
# after this PR:
#   lin.dtype == int64, lin == [8000040000]
#   NUMPY.unravel_index(lin, shape) → [[40000, 40000]]   (round-trips)

In real use this surfaces in phi.physics.fluid.make_incompressible on a 3-D Navier-Stokes grid as soon as c_dims.volume * u_dims.volume > 2**31 — for an N³ velocity-on-faces problem, that's around N ≥ 48.

What changed

• phiml/backend/_numpy_backend.py:362-378 — drop the unconditional .astype(multi_index.dtype), pick int64 when overflow would occur.
• tests/commit/backend/test__backend.py — new test_ravel_multi_index_int32_overflow regression case (large-shape round-trip + small-shape dtype preservation).

Testing done

• New regression test passes.
• tests/commit/backend/test__backend.py (43 cases) and tests/commit/math/test__sparse.py pass with the patch. One pre-existing failure on test_convert (JAX/DLPack GPU issue, unrelated to this PR) was confirmed to also fail on main.

Closes #15.