GPU tier: texture formats and anisotropic filtering (#2397)

[nicolas-maman]

Closes #2397.

Masks, normal maps, HDR images and compressed assets can now be sampled through the shared GPU shape, in contrib.vulkan, contrib.d3d12 and contrib.metal.

• Formats: texture_create_format(dev, w, h, format, mipmapped, linear, repeat, anisotropy) makes a 2D texture in R8, RG8, RGBA8 UNORM or sRGB, RGBA16F, RGBA32F, or BC1/BC3/BC4/BC5/BC7 (BC1, BC3 and BC7 also sRGB). texture_format_supported asks first. A block-compressed texture must be whole 4x4 blocks in every module, because Direct3D 12 requires it.
• Mip chains: a mipmapped uncompressed texture builds its chain on upload, in its own format.
  • Vulkan blits it, and refuses a format it cannot linear-blit.
  • D3D12 builds it on the CPU: bytes as bytes, half and float as floats, sRGB as linear light.
  • Metal uses generateMipmapsForTexture:, and refuses a 32-bit float chain where the GPU cannot filter one.
• Level uploads: texture_upload_level uploads a block format's chain, or one made offline, one level at a time. Only that level's layout or state changes. The texture binds once every level has pixels, and the refusal says which case it is.
• Anisotropy: the last argument filters anisotropically, up to the device's limit; texture_anisotropy reports what was granted.

Tests: each module's textures spec checks the following against pixels read back from an RGBA32F target:

• every uncompressed format, against known values, including sRGB 188 reading as linear 0.503 and half/float values outside 0..1;
• each format's chain at its 1x1 level: a 4x4 ring reads 0.75, where byte-averaged sRGB would read 0.52;
• known BC1, BC4 and BC7 blocks, decoded;
• a BC1 chain uploaded level by level, refused to bind until it is complete;
• 16:1 squeezed one-texel stripes, which blur to grey with the isotropic filter and stay sharp with the anisotropic one.

Run locally:

• Vulkan: NVIDIA under the validation layer with sync and best practices on, and lavapipe. Only the known small-dedicated-allocation advisories remain.
• D3D12: the GPU and WARP with the debug layer, 0 messages.
• Metal: cross-compiled for arm64 macOS; it first runs on macOS CI.

Follow-up filed: BC6H, ASTC/ETC2, signed formats, and cubes and arrays in other formats (#2402).