Speed up calculations

[olehz]

Speed up WMM calculations (~120x)

Why

wmmsub() re-read WMM.COF from disk, allocated and freed all model memory
on every single point. The Python wrapper looped point-by-point, creating
fresh ctypes objects and doing an os.chdir() round-trip per call (needed
because the C code opened the coefficient file by relative path).

For a 0.1° global grid (6.2M points) that meant 6.2M file reads.

Results

Single-threaded, per point:

	main	this PR
throughput	40.2 µs/pt	0.33 µs/pt

What changed

C (wmm_point_sub.c)

• New lifecycle: wmm_init(cof_path) / wmm_free(). Coefficients and scratch
  buffers are loaded once and reused.
• Batch entry points: wmm_eval_latlon_grid(), wmm_eval_grid(), wmm_eval_many().
• Legendre functions and (a/r)^(n+2) depend only on latitude and altitude, so
  grid evaluation hoists them out of the longitude loop; only the cos/sin(mλ)
  recurrence and the summation run per point.
• The Schmidt normalization ratios depend only on n/m, but MAG_PcupLow
  rebuilt them — with a malloc/free — on every point. They are now computed
  once at init.
• Dropped the secular-variation summation, which the Python API never exposed.
• wmmsub() is kept for backward compatibility.

Python (base.py)

• One ctypes call per array instead of one per point; numpy buffers are passed
  through without copying.
• Absolute path to WMM.COF, so the per-point os.chdir() hack is gone.
• Replaced assert in the hot path with real exceptions (assert is stripped
  under -O).
• transect() is now exported from the package.

Build

• Defaults to Release with -O3 (previously no build type was set at all, so
  the shared library was built unoptimized).
• Rebuilds automatically when the C sources are newer than the built library.

Breaking change

wmm() now returns a dict of numpy arrays instead of an xarray.Dataset:

mag = wmm2025.wmm(lat, lon, alt_km, yeardec)
mag["decl"]        # ndarray, was mag.decl.values
mag["glat"], mag["glon"], mag["time"]

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.