Add Bernstein–Khushalani fitter as an engine option in do_fit

[matthewholman]

This branch was developed with Claude Code, on top of an initial
implementation I had started.

Adds a Bernstein–Khushalani (BK) fitter as a second engine option in
do_fit. BK parameterizes the orbit in a tangent-plane (α/β/γ) frame
rather than barycentric Cartesian, which is much better conditioned
for distant short-arc targets where Cartesian-LM is near-degenerate.

do_fit gains an engine parameter:

• "cartesian" (default, unchanged behavior)
• "bk" (always use BK)
• "auto" (route on the Gauss IOD's heliocentric distance:
  r ≥ bk_threshold_AU → BK, else Cartesian)

The BK fitter is now distributed as a separate library, liborbfit:
https://github.com/matthewholman/liborbfit. It ships both the
liborbfit.so C library and a liborbfit Python package wrapping it
via ctypes. layup's src/layup/bk_orbitfit.py is a thin (~174-line)
adapter: it converts layup Observation objects to liborbfit's
Observation, calls liborbfit.do_bk_fit, and packs the
BKFitResult back into a layup FitResult — using liborbfit's
BKFitResult.to_cartesian() for the αβγ→ICRS-equatorial Cartesian
transform with analytic 6×6 covariance propagation.

Setup

Clone and build liborbfit, then export two env vars so layup can load
both the C library (via ctypes) and the Python wrapper:

git clone https://github.com/matthewholman/liborbfit
cd liborbfit
./setup_deps.sh && make
export LIBORBFIT_PATH=$PWD/liborbfit.so
export PYTHONPATH=$PWD/python:$PYTHONPATH

When the env vars are not set, the BK engine is unavailable: layup's
engine="auto" falls back to Cartesian, and explicit engine="bk"
calls return a FitResult with flag=3. Setup instructions also
live in src/layup/bk_orbitfit.py's module docstring.

Validation (diagnostic/scan, 98 cases)

• 87/88 cases succeed with all three engines (cartesian / bk / auto)
• Auto matches the per-case best of {BK, Cartesian} in 87.5% of
  cases (77/88). Routes to BK on 56, Cartesian on 35.
• Auto beats cartesian-only on 43 cases (distant short-arc) and
  beats BK-only on 16 cases (mainbelt / NEO-like).

tests/layup/test_bk_dispatch.py covers the dispatch logic; the
suite skips cleanly on machines without liborbfit installed, so CI
is unaffected.

Dependencies

Stacks on feat/pybind11-eigen-includes (PR #322 — adds
pybind11/eigen.h + pybind11/stl.h so Observation.rho_hat works
from Python). Once #322 lands, this PR collapses to its 3 new
commits.

Review Checklist for Source Code Changes

• Does pip install still work?
• Have you written a unit test for any new functions? — tests/layup/test_bk_dispatch.py
• Do all the units tests run successfully?
• Does Layup run successfully on a test set of input files/databases? — full diagnostic/scan
• Have you used black on the files you have updated? — formatting not strictly applied yet