test(core): gate u8 branch tests by backend types

[shinaoka]

Summary

• move the u8 literal branch runtime tests out of float test modules
• generate them only when a backend's unsigned-type list includes u8
• preserve one true/false test pair for CPU, CUDA, and SPIR-V while skipping unsupported WGSL/MSL paths

Root cause

The shared branch macro generated an Array<u8> kernel test once per floating-point type. WGSL has no supported u8 element, so the WGPU suite produced four guaranteed compiler panics.

Validation

• reproduced the original four WGPU failures
• focused WGPU filter: 0 tests, 0 failures after the change
• CPU and CUDA test lists each contain u8_ty::test_if_literal_u8_{true,false}
• cargo fmt --check
• cargo check -p t4a-cubecl-core
• cargo test -p t4a-cubecl-wgpu --lib: 375 passed, 14 ignored, 0 failed