pilot: Add offscreen 2D world renderer and Python binding (issue #754 PR2)

[tigercosmos]

!! BLOCK BY #899 !!

This builds on PR1's shared paint_world_2d routine to add an offscreen renderer that needs no widget and no running event loop, so a World can be turned into an image headlessly. The new C++ free function render_world_2d (in RPainter2d) fills the canvas background and draws the world geometry into a QImage; wrap_pilot.cpp exposes it to Python as pilot.render_world_2d, which returns PNG-encoded bytes, and pilot.save_world_2d, which writes a file whose format follows the filename suffix. Both raise ValueError on a None world or a non-positive size. This is the headless capture path the deterministic corpus and the PR3 pixel-test gate need, mirroring R3DWidget's saveImage/clipImage but for the strictly-2D widget.

This is PR2 of the issue-#754 stack and depends on PR1 (#899). Until PR1 merges to master, this PR's diff also includes PR1's commit (the paint_world_2d extraction); the relevant change here is the latest commit, "Add offscreen 2D world renderer and Python binding".

To use it, build a World, frame it with a 2D view transform, and ask for the image. The render functions paint with QPainter, so run them through the pilot binary, which provides the QApplication (plain python3 has none):

import modmesh
from modmesh import pilot

world = modmesh.WorldFp64()
world.add_circle(0.0, 0.0, 1.0)
world.add_rectangle(-2.0, -2.0, 2.0, 2.0)

view = modmesh.ViewTransform2dFp64()
view.pan(400.0, 300.0)   # world origin at the image center (width/2, height/2)
view.zoom = 60.0         # pixels per world unit

png_bytes = pilot.render_world_2d(world, view, 800, 600)   # -> PNG bytes
pilot.save_world_2d(world, view, 800, 600, "/tmp/world.png")

env PYTHONPATH=$(pwd) \
  ./build/rel<pyvminor>/cpp/binary/pilot/pilot.app/Contents/MacOS/pilot \
  --mode=python demo.py

The tests in test_pilot_2d.py render a known world headless and assert the result is a valid PNG of the requested size, that a world with geometry does not encode identically to an empty one (something is actually painted), that output is byte-stable across runs, and that the None-world and non-positive-size error paths raise. The endpoint-to-pixel mapping assertions remain PR3's gate.

Related to #754.

[yungyuc]

Leave inline annotation to indicate the parts that is not included in PR #899 and supposed to be reviewed.