refactor(glsl): accept BindingMap instead of hardcoded group*16+binding formula

[kolkov]

The GLSL writer currently uses group * 16 + binding to flatten WebGPU bindings to GL binding points. Rust naga accepts a BindingMap from the HAL which provides per-type sequential counters. This is more robust and matches the Rust wgpu ecosystem pattern. Depends on wgpu binding map refactor.

[kolkov]

Update: HLSL backend already has full BindingMap support (hlsl/backend.go:18-21). The GLSL backend only has *BindingBase offsets in Options.

The implementation should follow the HLSL pattern:

1. Add BindingMap map[ResourceBinding]BindTarget to glsl.Options
2. In flattenedBinding(), check BindingMap first, fall back to group*16+binding formula
3. wgpu#108 will then build BindingMap from pipeline layout and pass it

This issue blocks wgpu#108 (not depends on it — naga must add the API first, then wgpu uses it).