Modeling - Skip redundant optimization for flat deviations in GeomLib_CheckCurveOnSurface

[andymai]

Fixes #1372.

GeomLib_CheckCurveOnSurface runs, per sub-interval: a ~310-evaluation math_PSO swarm, a math_NewtonMinimum polish that is guaranteed to fail on flat deviation functions (singular Hessian), and a second full swarm as the failure fallback — even when the 3×N control-point seed sampling has already shown the deviation is zero to machine precision. On analytic geometry (line/circle on plane/cylinder/sphere/…, i.e. most prismatic CAD) this accounts for ~18% of the total cost of a boolean operation (callgrind numbers in #1372).

This change accepts the sampled deviation and skips the optimization pipeline when every control point lies below the geometric noise floor (Precision::SquareConfusion() on the squared distance):

• immediately, when the whole composition is analytic (3D curve and pcurve of type Line/Circle/Ellipse/Hyperbola/Parabola on an elementary surface): the deviation function of such a composition is a smooth low-degree function already resolved by the control-point sampling used to seed the swarm;
• after confirming flatness by resampling at half-step offsets (doubling the sampling density), for all other compositions.

Sub-interval handling, the seeding itself, and every non-flat path are unchanged: any sampled squared distance above the noise floor falls through to the existing PSO + Newton pipeline. Deviations that are real but flat (e.g. a pcurve on a cylinder of a different radius) are above the floor at every sample, so they take the existing path — covered by the new gtest.

Measurements (single-threaded; details in #1372):

• native (gcc 13 -O2): baseplate cut with 9 rounded-box tools 59.5 → 47.4 ms (−20.3%); 5-solid fuse 28.0 → 23.2 ms (−17.1%)
• WebAssembly (Emscripten 5.0.3, Node 24): same workloads 100.9 → 80.8 ms (−19.9%) and 51.2 → 41.4 ms (−19.1%)
• TargetFunc::Value evaluations: 495 728 → 304 832 with the analytic gate alone; the confirmation resample removes most of the remainder
• result shapes bit-identical on both workloads (topology counts, volume, area, tolerances, BRepCheck_Analyzer)

Validation:

• DRAW test groups boolean, blend, lowalgos: outcome sets identical between patched and unpatched builds on the same machine (2319 OK; identical pre-existing BAD/FAILED sets) — the change does not alter any test outcome
• GTest suite: no new failures; adds GeomLib_CheckCurveOnSurface_Test.cxx covering a coincident analytic composition (zero deviation), a deviating analytic composition (real constant deviation must be measured, not swallowed by the fast path), and line-on-plane

Especially relevant for single-threaded consumers (WebAssembly builds), where OSD_Parallel cannot hide the cost of this check.

[dpasukhi]

Dear @andymai

Thank you for your patch. To proceed with integration, please complete signing CLA process:
The links: Contribution Guide
Or https://dev.opencascade.org/get_involved

In case if you already have signed CLA, please share ID that OCCT team shared with your by email after accepting of CLA.

[andymai]

@dpasukhi CLA signed, just awaiting confirmation.