[Impeller] Metal backend crashes when binding a texture whose sampler was dead-code-eliminated #190034

Steps to reproduce

1. Write a Flutter GPU fragment shader that declares a sampler2D used only in code the optimizer removes, so SPIRV-Cross keeps it in the reflected resource list but drops it from the emitted Metal shader.
2. From Dart, bind a texture to that sampler with renderPass.bindTexture(shader.getUniformSlot('the_sampler'), texture).
3. Draw on Metal with flutter run -d macos --enable-flutter-gpu --enable-impeller.

Expected results

Binding a texture for a sampler the compiled shader does not use should be a harmless no-op, the way the OpenGLES and Vulkan backends already handle it. It should never crash the driver.

Actual results

The Metal backend passes the dead-code sentinel index uint32_t(-1) (4294967295) straight to setFragmentTexture:atIndex:, which has no range check, so the AGX driver crashes.

0  AGXMetalG15X  -[AGXG15XFamilyRenderContext setFragmentTexture:atIndex:]
1  FlutterMacOS  impeller::PassBindingsCacheMTL::SetTexture(impeller::ShaderStage, unsigned long long, id<MTLTexture>)
2  FlutterMacOS  impeller::RenderPassMTL::BindDynamicResource(impeller::ShaderStage, impeller::DescriptorType, impeller::SampledImageSlot const&, ...)

Code sample

uniform sampler2D tex_a;
uniform sampler2D tex_b;   // reflected, but dropped from the Metal shader
in vec2 v_uv;
out vec4 frag_color;
void main() {
  vec4 c = texture(tex_a, v_uv);
  if (false) { c += texture(tex_b, v_uv); }   // makes tex_b dead
  frag_color = c;
}

Then bind both tex_a and tex_b from Dart and draw. Binding tex_b crashes. (This shader shape needs a quick confirmation that it lands on the reflected-but-dropped path, see the note below.)

Root cause

impellerc still lists a dead sampler in its reflection but stamps that entry's index with uint32_t(-1), because get_automatic_msl_resource_binding returns ~0u for a sampler SPIRV-Cross removed from the emitted Metal. That value is written to the shader bundle, cached on the Dart Shader, and forwarded through RenderPassMTL::BindDynamicResource to setFragmentTexture:atIndex: with no range check. Every other backend avoids the problem. Vulkan binds by the SPIR-V binding decoration, and OpenGLES binds by name and skips a location of -1 (buffer_bindings_gles.cc, "The texture binding was optimized out of the shader. Continue.").

Proposed fix

Any one of these stops the crash, and they compose. None of them add per-draw work.

• Skip a texture binding whose ext_res_0 is uint32_t(-1) when loading the shader in lib/gpu/shader_library.cc, so the bind becomes a no-op like OpenGLES. This works with already-built shader bundles.
• Stop emitting sampled_images entries with a sentinel index in reflector.cc, so the bundle reflection matches the compiled shader.
• As a backstop, skip a bind whose index is the sentinel or out of range in pass_bindings_cache_mtl, so no invalid binding can ever reach the driver.