fix(troika-three-utils): avoid inout params in vertex transform so text renders in Safari/ANGLE Metal

[rokas-th]

Fixes #390.

Problem

createDerivedMaterial emits the vertex transform as a function taking inout parameters, called with the hoisted shader globals:

void troikaVertexTransform${key}(inout vec3 position, inout vec3 normal, inout vec2 uv) {
  ...
}
...
troikaVertexTransform${key}(troika_position_${key}, troika_normal_${key}, troika_uv_${key});

On Safari (macOS, ANGLE Metal backend), ANGLE hoists those globals into a __metal_generic-address-space struct, then wraps each inout argument in an ANGLE_InOut<T> helper whose conversion operator returns thread T&. The MSL compiler rejects binding the __metal_generic global to a thread reference, so every derived material fails to link:

THREE.WebGLProgram: Shader Error 1282 - VALIDATE_STATUS false
Program Info Log: Internal error while linking shader. MSL compilation error:
  error: reference to type 'thread float3' could not bind to an lvalue of type '__metal_generic float3'
      operator thread T &() { return mTemp; }

followed by an endless stream of WebGL: INVALID_OPERATION: useProgram: program not valid. The result is that no troika-three-text renders in Safari. Chrome/Firefox are unaffected because their ANGLE builds use different backends. This is the ANGLE translator defect tracked upstream in WebKit bug 250169 (see also 319958) — the underlying bug is Safari's, but troika can sidestep it entirely.

Fix

Emit the transform as a parameterless function with explicit copy-in / copy-out locals instead of inout parameters. It's semantically identical to the previous version (same local names, same read-then-write-back semantics), but never passes a global by reference, so it compiles cleanly under ANGLE Metal.

Verification

Running this as a patch-package patch against 0.52.4 in production: text renders correctly again in Safari, with visually identical output in Chrome and Firefox.

---

Reported and fixed while debugging this in our app; PR opened at maintainer's request (per #390) so the fix is attributed correctly.

[lojjic]

Thank you!