Python-side dimension-limit checks, large-dimension tests, and 1-column igd support (#72)

[blankjul]

Closes the remaining checklist items from #72.

Commit 1 — Python-side dimension-limit checks + tests

Adds a _check_nobj() helper and per-function ValueErrors so that inputs beyond what the C library supports are rejected in Python instead of reaching C as undefined behavior:

Functions	Limit	Before this PR
igd, igd_plus, avg_hausdorff_dist, epsilon_additive, epsilon_mult, pareto_rank, is_nondominated, any_dominated	255 (range of dimension_t)	256 columns segfaulted (dimension_t wraparound)
hypervolume (function + class), hv_contributions, hv_approx	31	32 columns silently returned a wrong value (release build)
epsilon_additive, epsilon_mult	at least 2	1 column read out of bounds and returned garbage (see below)

New test_dimension_limits.py:

• large-dimension correctness at 32/64/128/255 columns for igd, igd_plus, epsilon_additive, pareto_rank, is_nondominated — verified against independent pure-numpy references (~0.3 s total),
• error-raising tests at each limit (at-limit succeeds, limit+1 raises).

Note on epsilon: while testing the lower bound we found that epsilon_helper_ is unrolled over the first two dimensions (MAX(eps_value_agree_(0), eps_value_agree_(1)) unconditionally), so a 1-column input performs an out-of-bounds read and returns garbage — e.g. epsilon_additive returned 0.616 where the correct value is 0.009 (formula cross-checked against moocore at 2–3 columns). Reachable from Python in all released versions; now rejected with a clear error.

Commit 2 — relax igd.h to ASSUME(1 <= dim) (drop if you disagree)

The gd_common_helper_ loop is fully general in dim, so the 2 <= dim lower bound is purely conservative there: with 1 <= dim, igd/igd_plus/avg_hausdorff_dist compute exact results for 1-column input (tests included in test_dim1_metrics.py) and higher dimensions are unchanged.

Why legalize rather than reject: pymoo 0.6.2 (current release) calls moocore.igd on design-space data, so any single-variable problem passes a 1-column array in production. Today that violates the ASSUME (it happens to work on GCC release builds but would assert under DEBUG=1). A ValueError here would retroactively break those installs; relaxing the assume makes the existing behavior well-defined. epsilon.h is deliberately untouched — its 2 <= dim is load-bearing (see above).

This is a separate commit so it's easy to drop if you'd rather handle the lower bound differently.

Verification

• New tests: 44 (test_dimension_limits.py 40, test_dim1_metrics.py 4)
• Full suite on this branch: 115 passed, 2 xfailed (Linux/GCC, Python 3.11)

Open questions (not in this PR)

• Would you prefer the limit constants exposed from C (via cffi) instead of mirrored in _moocore.py? Happy to change.
• NaN/inf checking (your earlier point): confirmed igd with NaN silently returns 0.0 today. A SciPy-style opt-out check_finite=True kwarg would avoid always paying the extra pass — can do as a follow-up if you want it.
• Numpy fallback above 255: skipped for now per your comment; can contribute later if demand appears.

Downstream context: anyoptimization/pymoo#793.