fix(cpp): lower CUDA complex casts with cuComplex helpers

[GiggleLiu]

Observation

CUDA compilation fails when a scalar initializes a complex value. Zero fill, for example, generated:

cuDoubleComplex(uint32(0))

NVRTC rejects this because CUDA complex types are C structs without converting constructors. The same lowering also breaks scalar/complex and cf32/cf64 conversions.

Issue

Assign::format_scalar emitted TargetType(input) for conversions that involve CUDA complex types. Complex conversions must construct or extract components explicitly. Complex-to-bool also inspected only the real component, so 0 + 1i became false and disagreed with constant folding.

Fix

• construct complex values with make_cuFloatComplex or make_cuDoubleComplex
• convert complex precision component by component
• extract the real component for numeric scalar casts
• treat a complex bool as true when either component is nonzero, in codegen and constant folding
• add regression tests for each conversion path

The PR also includes two mechanical fixes required by the current Clippy pass.

Verification

• cargo test -p t4a-cubecl-cpp --lib
• cargo test -p t4a-cubecl-ir --lib
• cargo test -p t4a-cubecl-core --lib
• cargo clippy --no-deps -- --deny warnings
• cargo fmt --all -- --check
• NVIDIA A800, CUDA 12.1: 45 existing complex tests passed
• NVIDIA A800, CUDA 12.1: focused scalar, precision, real, and bool cast tests passed