Add UV texturing: checker and image textures with bilinear filtering

[erichanwang]

Why

Material only rendered flat color or the isGrid special case (gridColor1/2,
gridSize) bolted directly onto Material - a one-off that couldn't be
extended to anything else. This replaces it with a proper Texture
abstraction Material can hold.

What changed

• Sphere and Plane now compute UV on every hit (standard spherical
  parameterization for Sphere; an orthonormal tangent basis for Plane).
  Block gets box-mapped UV nearly for free from its existing per-face
  normal logic.
• Material::texture (a Texture*) replaces isGrid/gridColor1/
  gridColor2/gridSize entirely.
• CheckerTexture: the procedural checker, migrated from the old
  world-space formula onto UV.
• ImageTexture: loads a PPM (P3 ascii or P6 binary) and samples it with
  bilinear filtering, wrapping UV instead of clamping.
• main.cpp and main_control_demo.cpp migrated off isGrid onto
  CheckerTexture(gridColor1, gridColor2, 1.0) - same checker pattern,
  verified pixel-identical by eye against the prior render.

Two things that looked right and weren't (both fixed, both covered by new tests)

1. Render cost of the texture pointer. Material (and the HitRecord
   holding it) gets copied by value on every ray-object test throughout
   Scene/BVH traversal. A first pass held the texture in a std::shared_ptr,
   which looked like the obviously safe choice - but its atomic refcount
   bump on every one of those copies cost real time: measured 0.8s vs 0.36s
   (~2.2x) rendering main.cpp's scene with vs without a textured ground
   plane, on otherwise identical code. Switched to a raw Texture*, matching
   this codebase's existing style of never freeing Object* either (see
   Scene::addObject) - textures live for the process lifetime, same as the
   objects referencing them. Re-measured afterward: ~0.5s both ways,
   texture path sometimes faster, i.e. within noise.

2. Plane UV sign convention. The first tangent-basis choice reproduced
   the old floor(point/gridSize) checker parity correctly for generic
   points, but silently flipped phase for any point sitting exactly on a
   grid-line boundary (an integer world coordinate) - the sign-flip trick
   it relied on only preserves floor() parity for non-integer inputs, not
   at the boundary. Fixed by picking axes so the ground plane's UV equals
   its world x,z exactly (no sign flip at all), which matches the old
   formula bit-for-bit including at boundaries.

Tests added (tests/test_math_physics.cpp)

• testSphereUVMapping: closed-form UV at both poles and four equator
  points (u,v are exactly computable there).
• testPlaneCheckerMatchesOldGridFormula: checks the new UV-driven
  checker against the old world-space formula across a grid of points,
  deliberately including exact grid-line boundaries (this is the
  regression test for pitfall Replace O(n^2) broad phase with a spatial hash; fix box collision #2 above).
• testBilinearImageTexture: hand-computed interpolation over a 2x2 PPM
  fixture written to build/ and read back - covers the center point
  (equal-weighted average of all 4 texels) and an edge-midpoint (pure
  horizontal blend, zero vertical blend).

All 18 tests pass. src/main.cpp, main_blocks.cpp, main_parachutes.cpp,
and main_control_demo.cpp all build clean under
-std=c++17 -Wall -Wextra -Werror -O2 -pthread, as does the test binary.
Rendered src/main.cpp and inspected the output PPM - ground plane checker
is visually identical to before, and it's correctly visible reflected in
the chrome and glass spheres.

What to check

• The two commit-message pitfalls above are exactly what a reviewer should
  scrutinize: the raw-pointer ownership tradeoff (fine here, given the
  codebase's existing style, but worth a second opinion), and the boundary
  case in testPlaneCheckerMatchesOldGridFormula.
• bench/benchmark.cpp still references the old isGrid/gridColor1/2/
  gridSize fields and will fail to build until it's migrated onto
  Texture - left untouched here since another change is in flight
  against bench/.
• Not attempted: normal/bump mapping (the suggested follow-on if time
  allowed) - ran out of scope/time budget for this pass.