pyglet 3.0.dev6 migration

[pvcraven]

Summary

Migrate Arcade's pinned pyglet dependency directly from 3.0.dev3 to 3.0.dev6 (no dev4 intermediate). This is a hard cutover: pyglet dev5+ removed
window._matrices and restructured how window view/projection/viewport are stored, which broke Arcade's rendering pipeline in three distinct ways. All three are fixed
here, verified empirically against the real pyglet dev6 source and a full local test run on native Windows.

Background

Between pyglet dev4 and dev5, pyglet:

• Removed window._matrices (and its .ubo) entirely.
• Moved window view/projection/viewport onto a lazily-created default_camera (pyglet.window.camera.camera2d.Camera2D), backed by a per-frame ring-buffer UBO.

A naive version bump crashes immediately (AttributeError: 'Window' object has no attribute '_matrices'), and would go on to cause unbounded UBO growth once past
that, because Arcade drives its own render loop and doesn't participate in pyglet's frame-resource lifecycle.

What changed

1. Arcade now owns its window-block UBO directly (arcade/context.py)

ArcadeContext allocates a single fixed-size (128-byte) buffer in __init__ instead of reusing window._matrices.ubo. projection_matrix/view_matrix setters
write straight into this buffer. This buffer's identity never changes for the lifetime of the context — no ring buffer, no reallocation, so growth is impossible by
construction rather than by careful bookkeeping.

OpenGLArcadeContext/WebGLArcadeContext.bind_window_block() (arcade/gl/backends/{opengl,webgl}/context.py) now bind this buffer via the existing
Buffer.bind_to_uniform_block() method instead of reading self._window_block.buffer.id from pyglet.

2. Camera write-path fix (arcade/camera/{orthographic,perspective,camera_2d}.py)

Found via failing camera unit tests, not anticipated in the original migration plan: OrthographicProjector.use(), PerspectiveProjector.use(), and Arcade's own
Camera2D.use() were writing matrices through self._window.projection = ... / self._window.view = ... — i.e. through pyglet's window properties, not Arcade's.
Under dev3 this was harmless (both paths pointed at the same underlying buffer). Under dev6, pyglet's properties write into its own ring buffer — a completely
different buffer from the one Arcade now owns — so camera changes silently never reached the GPU. Fixed by routing all three through
ctx.projection_matrix/ctx.view_matrix instead, matching the pattern DefaultProjector already used.

3. DefaultProjector pyglet-compatibility shim (arcade/camera/default.py)

Found via the example/tutorial integration tests: pyglet's own batch-draw pipeline (used internally by pyglet.text.Label.draw(), which arcade.Text delegates to)
falls back to window.default_camera when no explicit camera is given. Since Arcade's Window.default_camera property override intercepts every access to that
name, pyglet's fallback gets Arcade's DefaultProjector instead of its own Camera2D, and calls methods on it that only a real pyglet camera implements — crashing
with AttributeError.

Rather than reimplementing pyglet's internal camera/view-hierarchy protocol (which would tightly couple Arcade to pyglet internals — exactly what this fix is trying to
avoid), DefaultProjector gained a minimal duck-typing shim:

• .view → returns self
• .begin() → no-op (preserve whatever camera state Arcade already has bound)
• .get_group_scissor_area() → always None
• .projection / .view_matrix (get/set) → thin aliases over existing state

This also fixes window.projection/window.view (pyglet's base properties, which delegate to self.default_camera.projection/.view_matrix) for any test
infrastructure or downstream code that reads them directly.

New test infrastructure

• tests/unit/rendering/ — a reference-image comparison harness (scenes.py, image_compare.py, generate_baseline.py, test_dev6_reference_images.py) with 4
  checked-in baseline PNGs captured on dev3 before the pin change. No such harness existed before this PR.
• tests/integration/test_ubo_stress.py — 100-cycle window/draw/close and ctx.reset() stress tests asserting zero "Growing UniformBufferObject" warnings and a
  stable buffer identity.

Testing performed

• Camera/window unit suite: 125/125 pass
• Reference-image comparison: 4/4 scenes pixel-identical to dev3 baselines
• UBO stress tests: 100 window cycles + 100 reset cycles, zero growth warnings
• Full suite (pytest tests/) on native Windows: 1283 passed, 10 failed/1 error — confirmed byte-identical to running the exact same suite against the untouched
  dev3 install on the same machine. These pre-existing failures (Windows DPI-scaling assumptions in one test, shared-global-window test-order flakiness in a few
  others, an unrelated encoding bug in a docstring checker) are unrelated to pyglet version and out of scope for this PR.
• .github/workflows/test.yml (Linux + xvfb CI) is untouched — verified via git diff.

Breaking changes

• Exact pin pyglet==3.0.dev6 (was 3.0.dev3); no dual code paths, dev3/dev4 no longer supported.
• Internal only: window._matrices no longer exists (removed by pyglet). Advanced users mixing raw pyglet camera/window code with Arcade should note
  window.default_camera is now pyglet's own Camera2D, distinct from arcade.Window.default_camera (Arcade's DefaultProjector) — both exist simultaneously and
  serve different roles. See CHANGELOG.md.
• No change to Arcade's public API surface (camera classes, ArcadeContext properties, Window.default_camera, get_image()/get_pixel()) — verified by the full
  test suite.

Test plan (for reviewers)

• uv sync --no-group docs resolves and installs cleanly
• uv run pytest tests/unit/camera tests/unit/window -v — all pass
• uv run pytest tests/unit/rendering -v — all 4 reference-image comparisons pass
• uv run pytest tests/integration/test_ubo_stress.py -v — both stress tests pass
• uv run pytest tests/integration/examples tests/integration/tutorials -v — no AttributeError from default_camera
• Existing Linux+xvfb CI stays green