TextureSampleLevel i32/u32 overloads on sampled textures exceed WGSL spec; add textureGather for cube types

## Two follow-up items from PR #156

### 1. `TextureSampleLevel<_, i32>` and `<_, u32>` on non-depth sampled textures

Per [WGSL spec §17.7.13](https://www.w3.org/TR/WGSL/#texturesamplelevel), the `level` parameter for `textureSampleLevel` must be **`f32`** for sampled textures (`texture_1d`, `texture_2d`, `texture_2d_array`, `texture_3d`, `texture_cube`, `texture_cube_array`). Only **depth** textures accept `i32`/`u32` levels.

The codebase implements `TextureSampleLevel<_, i32>` and `TextureSampleLevel<_, u32>` for all of these, which means code that compiles in Rust world would fail WGSL validation. This is a pre-existing issue (already present on `Texture2D` before #156) and was propagated to the new types for consistency.

Affected (from #156): `Texture1D` (lines 369, 381), `Texture3D` (lines 1340, 1359), `TextureCube` (lines 1584, 1603).

### 2. `textureGather` for `TextureCube` and `TextureCubeArray`

Per [WGSL spec §17.7.2](https://www.w3.org/TR/WGSL/#texturegather), `textureGather` has overloads for `texture_cube` and `texture_cube_array` (with `ST` in {f32, i32, u32}). PR #156 implemented the `textureSample` family for cube types but did not add the `TextureGather` / `TextureGatherArray` trait impls.

### Proposed fix

- Remove the `i32`/`u32` level overloads from all **non-depth** sampled texture types (`Texture1D`, `Texture2D`, `Texture2DArray`, `Texture3D`, `TextureCube`, `TextureCubeArray`), keeping only `f32` for those. Depth textures keep `i32`/`u32`/`f32`.
- Add `TextureGather` impls for `TextureCube` and `TextureGatherArray` impls for `TextureCubeArray` with `u32`/`i32` array indices, including the CPU-side face selection and 4-texel gather logic.

Start by locating the TextureSampleLevel implementations for Texture1D, Texture2D, Texture2DArray, Texture3D, TextureCube, and TextureCubeArray, then compare their overloads with the WGSL textureSampleLevel rules. Next locate the TextureGather and TextureGatherArray trait implementations and the cube texture code added by PR #156. Done means non-depth sampled textures retain only f32 level overloads, depth textures retain their permitted levels, and cube and cube-array gather implementations cover f32, i32, and u32 types with the specified indices and CPU-side gathering.

Written by the indexing model from the issue text.