[SYCL] Trim builtins header dependencies for compile time

[koparasy]

Summary

Reduce frontend compile-time overhead in sycl/include/sycl/detail/builtins/builtins.hpp by narrowing its dependencies and excluding host-only helpers from device compilation.

This change:

Replaces the sycl/detail/vector_convert.hpp inclusion with the narrower headers actually needed here: sycl/detail/generic_type_traits.hpp and sycl/half_type.hpp.

Wraps:

1. builtin_default_host_impl and
2. builtin_delegate_to_scalar

in #ifndef __SYCL_DEVICE_ONLY__ so they are not parsed during device-only compilation.
The intent is to reduce parsing and template-instantiation cost for translation units that pull in sycl/builtins.hpp, while preserving existing behavior.

Motivation

Compilation-time tracing for a device-side SYCL workload that includes sycl/builtins.hpp showed that the cost is frontend-dominated. This header is on a hot include path, so avoiding an unnecessary transitive dependency and skipping host-only helper code in device mode reduces frontend work in the common device compilation path.

Rough numbers:

HEAD compiler: 
  iter 1: frontend=1183.220 ms  backend=44.570 ms    execute=1234.452 ms  trace=/tmp/sycl-trace-rins/orig/orig.iter1.json
  iter 2: frontend=1178.349 ms  backend=44.313 ms    execute=1229.331 ms  trace=/tmp/sycl-trace-rins/orig/orig.iter2.json
  iter 3: frontend=1186.436 ms  backend=45.257 ms    execute=1238.330 ms  trace=/tmp/sycl-trace-rins/orig/orig.iter3.json
  iter 4: frontend=1180.083 ms  backend=44.657 ms    execute=1231.393 ms  trace=/tmp/sycl-trace-rins/orig/orig.iter4.json
  iter 5: frontend=1176.515 ms  backend=44.306 ms    execute=1227.457 ms  trace=/tmp/sycl-trace-rins/orig/orig.iter5.json
  avg:     frontend=1180.921 ms  backend=44.621 ms    execute=1232.193 ms

This Branch:
  iter 1: frontend=1087.281 ms  backend=45.098 ms    execute=1138.897 ms  trace=/tmp/sycl-trace-rins/new/new.iter1.json
  iter 2: frontend=1088.384 ms  backend=44.814 ms    execute=1139.661 ms  trace=/tmp/sycl-trace-rins/new/new.iter2.json
  iter 3: frontend=1087.635 ms  backend=44.241 ms    execute=1138.346 ms  trace=/tmp/sycl-trace-rins/new/new.iter3.json
  iter 4: frontend=1086.537 ms  backend=44.334 ms    execute=1137.336 ms  trace=/tmp/sycl-trace-rins/new/new.iter4.json
  iter 5: frontend=1091.378 ms  backend=44.172 ms    execute=1142.013 ms  trace=/tmp/sycl-trace-rins/new/new.iter5.json
  avg:     frontend=1088.243 ms  backend=44.532 ms    execute=1139.251 ms

when compiling the benchmark (gemm_sycl.cpp) provided in JIRA ticket (CMPLRLLVM-73941). This is roughly a 7% compile time reduction.