[SYCL] Add scalar-only builtins entry heade

[koparasy]

This PR adds sycl/builtins_scalar.hpp, a lighter-weight entry point for code that uses only scalar builtin functions.

Today, scalar-only users still pull in the full gentype/vector/marray builtin machinery through sycl/builtins.hpp. This change refactors the builtin implementation so scalar-safe pieces are separated from gentype-specific utilities, while keeping the existing full builtins path unchanged.

The motivation is compile-time reduction for scalar-only consumers. In a minimal softmax kernel where the only source difference is including sycl/builtins.hpp vs sycl/builtins_scalar.hpp, device-only syntax checking time dropped from 0.632s to 0.310s:

> diff softmax_kernel.cpp softmax_kernel_scalar.cpp 
6c6
< #include <sycl/builtins.hpp>
---
> #include <sycl/builtins_scalar.hpp>

> time clang++ -fsycl -fsycl-device-only -ftime-trace   softmax_kernel.cpp -fsyntax-only 
real    0m0.632s
user    0m0.587s
sys     0m0.038s
> time clang++ -fsycl -fsycl-device-only -ftime-trace   softmax_kernel_scalar.cpp -fsyntax-only 

real    0m0.310s
user    0m0.280s
sys     0m0.030s