Return references from Id lookups, and remove the glad dependency

[keithlostracco]

Closes #52. Closes #53.

#52 — channel(Id) returned a pointer that could never be null

Both Id overloads resolve through unordered_map::at, which throws on a
miss, so the only outcomes were a valid non-null pointer or an exception. The
Channel* return invited null checks that are dead code, and it was the odd
one out — the index and name overloads return references.

All four (channel(Id) / operator[](Id), const and non-const) now return
references. Miss behavior is unchanged: still std::out_of_range.

Breaking for callers, mechanically: -> becomes ..

The test updates deliberately keep &a == &b identity comparisons rather than
rewriting to a == b — Channel::operator== is value equality, so the
simpler-looking form would have silently weakened 49 assertions from identity
to equality.

#53 — glad removal, and the CMake 4 workaround goes with it

glad 0.1.36 is the last 0.1.x release and declares
cmake_minimum_required(VERSION 3.0), which CMake 4 rejects — hence the
CMAKE_POLICY_VERSION_MINIMUM 3.5 workaround. glad 2.x is not a drop-in: no
root CMakeLists.txt, and it needs Python at configure time.

Removing glad was simpler than migrating to it:

• ImGui already bundles imgui_impl_opengl3_loader.h and uses it when no
  loader macro is defined — the IMGUI_IMPL_OPENGL_LOADER_GLAD definition was
  the only thing redirecting it to glad
• the sole direct GL calls in curve_visualization are glViewport,
  glClearColor and glClear, all OpenGL 1.1 core, available from the GL
  header GLFW includes and already linked via OpenGL::GL

One dependency and one workaround removed together.

Verification

Against CMake 4.4 with no policy override: configure, build of library +
tests + both examples, and 70/70 tests pass.

The viewer was launched from a build with glad and a build without, each
from a clean layout. Both write an identical imgui.ini:

Window	with glad	without glad
###CurvesPlot	Pos=60,60 Size=32,39	Pos=60,60 Size=32,39
Curve Editor	Pos=60,60 Size=156,116	Pos=60,60 Size=156,116

So the change is behavior-neutral. Note this also documents a pre-existing
quirk, present on both sides and not addressed here: on a first run with no
saved layout the plot window auto-sizes to 32x39 and sits behind the Curve
Editor, so a new user sees no plot until they resize or dock it. Worth a
separate issue.

Also

imgui.ini is written into the working directory by the example and was not
ignored, so it appeared as untracked noise in the repo root. Now ignored, along
with build-*/.

[keithlostracco]

Added: the example rendered no plot on a clean checkout

Investigated whether this branch caused it. It did not — but it is a real bug
and it is fixed here.

Root cause. Curves Plot was opened with no size, so ImGui auto-fitted it
to its content. That content is ImPlot::BeginPlot(..., ImVec2(-1,-1), ...) —
"fill the available space". On the first frame neither has a size to work from,
so both resolve to nothing, the window collapses to 32x39, and it lands behind
Curve Editor, which defaults to the same position. ImGui then persists that
size to imgui.ini, so it stays collapsed on every later run.

Not a regression. A build of bd5f9f9 — before anything in this branch —
writes an identical layout:

Build	###CurvesPlot
bd5f9f9 (pre-branch)	Pos=60,60 Size=32,39
with glad (post-#50)	Pos=60,60 Size=32,39
without glad	Pos=60,60 Size=32,39

It was only ever masked by a saved imgui.ini from having dragged the window
out once.

Fix. Both windows get a first-run position and size derived from the
viewport work area, laid out side by side. ImGuiCond_FirstUseEver means an
arranged layout is still restored from imgui.ini as before.

After the fix, a clean first run gives:

[Window][###CurvesPlot]   Pos=12,12      Size=2460,1256
[Window][Curve Editor]    Pos=2484,12    Size=704,1256

Verified visually: the plot renders with axes, all five curves, and their
keyframes and tangent handles.