simplifier: Use double precision for quadric storage and evaluation

[snowbldr]

meshopt_simplify has a position-dependent floor on target_error: geometry
far from the mesh bounding-box minimum cannot be simplified below an error of
roughly 3e-4 × extent, while identical geometry near the bbox minimum
simplifies to completion.

Cause

The expanded quadric form r = vᵀAv + 2bᵀv + c cancels catastrophically in
single precision. For a plane with offset d (after internal rescale),
quadricFromPlane computes b = d·w and c = d²·w, whose independent
float32 roundings don't cancel — evaluating at a point on the plane
leaves a residual r ≈ 1e-7·w·d² instead of 0, i.e. a distance-error noise
floor of ~3e-4·extent·d. Since the internal rescale anchors at the bbox
minimum, d — and the floor — grows with distance from the min corner.
Collapse candidates on exactly flat, exactly coplanar surfaces then rank
above tight error limits and the region never simplifies.

Repro

Two identical 150×150 flat plates, one at z=0 and one at z=900, in a mesh of
extent 1000 (the offset must be along the plate normal, and the far
plate's rescaled offset must be inexact in binary — z=0.9 here; a tangential
offset leaves b = c = 0 exactly and hides the effect). Simplify with
target_index_count = index_count/100, target_error = 5e-5:

	master (float)	this PR (double)
plate A (z=0)	2 tris	446 tris
plate B (z=900)	38,602 tris	454 tris
total vs target	38,604 / 900	900 / 900
reported error	4.75e-05 (noise floor)	2.56e-06 (true geometric)

Master, unable to make progress on plate B, spends the entire reduction
budget over-collapsing plate A; this PR reaches the requested target and
stops. Asked for the actual minimum (target = 4 triangles), master still
returns 38,604 while this PR returns exactly 4 — 2 per plate.

Result meshes rendered with edges (master left/right = near/far plate, then
this PR):

Found in the wild on a 127mm CAD part where one exactly-planar box wall kept
~17M render-resolution triangles at tight target_error while the
geometrically identical opposite wall collapsed to a handful; with this
change the whole part reaches 16k triangles with measured surface deviation
of 2e-4 mm.

Change

Confined to the Quadric/QuadricGrad fields, the products in
quadricFromPlane, and the accumulation in quadricEval/quadricError
(22 insertions, 19 deletions). Storing double without computing the b/c
products in double is not sufficient — the coefficients are already rounded
at construction. Plane fitting from vertex positions stays in single
precision; its per-triangle rounding is second-order for this effect.

Includes a regression test (simplifyPrecision in demo/tests.cpp) that
builds the two-plate geometry procedurally and asserts symmetric collapse —
it fails on master and passes with this change.

make test passes; builds warning-free with -Wall -Wextra -Wshadow.

Trade-off

This doubles quadric array memory (per remap-unique vertex) and may
measurably affect simplification throughput on some targets. If that's
unacceptable for game-LOD workloads — which never request errors below the
noise floor and are unaffected by the bug — the same change works gated
behind a compile-time define, or I'm happy to rework it in whatever
direction you prefer (recentring the internal rescale only halves the worst
case, so some precision change appears necessary for a full fix).