[codex] Prepare t4a CubeCL packages

[shinaoka]

Summary

• Rename the minimal tenferro-required CubeCL publish closure to t4a-cubecl* packages while preserving Rust crate import names.
• Mark the packages and README text as a temporary tensor4all fork until the required CubeCL patches are upstreamed.
• Replace the publish workflow with the t4a-* publish graph and exclude unused CPU/HIP/SPIR-V packages from the minimal release path.

Notes

• This branch is intentionally based on the CubeCL commit currently pinned by tenferro-rs: f5e5ec178f9aebca9362b829ffef708f720ff692.
• Package versions are 0.10.0; downstream crates should use Cargo package aliases such as cubecl = { package = "t4a-cubecl", version = "=0.10.0" }.

Verification

• cargo fmt --all --check
• cargo check -p t4a-cubecl --no-default-features --features 'std,stdlib,wgpu'
• cargo check -p t4a-cubecl --no-default-features --features 'std,stdlib,cuda'
• cargo publish --dry-run --no-verify --allow-dirty -p t4a-cubecl-zspace
• cargo publish --dry-run --no-verify --allow-dirty -p t4a-cubecl-common
• cargo publish --dry-run --no-verify --allow-dirty -p t4a-cubecl-macros-internal
• cargo package --list --allow-dirty for all 13 t4a-cubecl* packages
• cargo metadata --format-version 1 confirmed no CubeCL git dependency leakage