Use Eriksson's formula for acute triangle areas

[jagrutipatilp]

Addresses #190.

Summary

• Use Eriksson's one-atan2 spherical excess formula when all three triangle edges are at most 90 degrees.
• Compute the determinant from (b - a) x (c - a) to retain relative accuracy for small triangles.
• Keep the existing l'Huilier/Girard implementation for triangles with an obtuse or near-antipodal edge, where Eriksson's denominator can become ill-conditioned.
• Update the existing near-degenerate regression to check its 100-digit reference area rather than accepting zero.
• Add acute and fallback-path microbenchmarks.

Numerical validation

A 100,000-triangle comparison against a 100-decimal-digit oracle covered well-shaped, tiny, randomly rotated skinny, and near-90-degree triangles. The largest observed absolute candidate error was 8.88e-16, below the documented 5e-15 maximum error.

An unguarded formula was also tested on near-antipodal inputs and produced errors as large as 0.65 steradians. The pairwise non-negative dot-product guard avoids that unstable region and leaves those triangles on the existing implementation.

Benchmark results

Release build, LLVM 23.1.0, x86-64 Windows; five repetitions:

Input	Existing median CPU	This change median CPU	Ratio
Small acute	140 ns	27.6 ns	5.1x faster
Large acute	160 ns	43.6 ns	3.7x faster
Obtuse fallback	267 ns	302 ns	1.13x slower

The fallback overhead is the cost of the three pairwise dot-product checks.

Tests

• s2measures_test
• s2loop_measures_test
• s2shape_measures_test
• s2shape_index_measures_test
• Full project compilation
• 122 of 123 test executables passed locally. The remaining unrelated ExactFloatTest.frexp case expects glibc's NaN/Infinity exponent behavior and fails with the Windows MinGW C runtime.