[SYCL] Remove iostream from the SYCL header path (#22047)
## Motivation
Downstream kernel-compilation flows are moving to a model where the
device pass compiles against a reduced C++ standard library (LLVM libc++
plus a small vendor libc subset). In that model
`<iostream>`/`<ostream>`/`<istream>` are unsupportable: they declare
host-only globals and emit per-TU static initializers
(`std::ios_base::Init`) that have no meaning in device JIT/AOT
pipelines. Today SYCL public headers pull them transitively into every
device compilation, even with `-fsycl-device-only`.
This PR removes the dependency structurally, superseding the
macro-gating approach of #22236.
## The contract
Every header under `<sycl/khr/...>` is now guaranteed free of iostream
headers and the `iostream_proxy` shim in both compilation passes. New
lit test `device_only_no_iostream.cpp` locks this: it compiles an
umbrella of all KHR headers and asserts no iostream pull via the `-MD`
dependency listing.
## Approach
Three orthogonal fixes, one per leak mechanism:
1. **Stream operators → runtime.** The five inline stream operators in
public headers (`backend`, `half` ×2, `bfloat16` ×2) now live in
`sycl/source/printers.cpp` as `__SYCL_EXPORT` symbols; headers keep
forward declarations plus `<iosfwd>`. Four ESIMD type-traits headers had
identical operators with zero test coverage — removed.
2. **Inline diagnostic code → runtime functions.**
`__SYCL_REPORT_EXCEPTION_TO_STREAM`, `defaultAsyncHandler`, and the
`Adjust` lambda in `range_rounding.hpp` no longer write to streams from
header code; each now calls an exported runtime function.
3. **`iostream_proxy.hpp` → runtime-private.** Moved from the public
include path to `sycl/source/detail/`. While there, the 7 runtime TUs
including `<iostream>` directly were switched to the proxy: zero
`<iostream>` includes left in the runtime, single `__ioinit` static,
smaller `.so`.
**ABI:** 8 new exported symbols (the 5 operators + 3 diagnostic
functions), additions only — previously inline, never exported. Added to
`sycl_symbols_linux.dump`.
**Compile time** (via `measure_split_headers_compile_time.py`):
aggregate **−4.9%** across all 39 KHR split headers; standouts
`half.hpp` −60.3%, `exception.hpp` −21.9%. Regressions are <2% and
within noise.
## Reviewing this PR
The diff stat (~725 files) overstates the review surface. The real
change is **~10-15 files** (header restructuring, `printers.cpp`, the
proxy move, the new test). The other ~660 are mechanical one-line
`+#include <iostream>` adds in e2e tests that were free-riding on the
transitive pull — include-what-you-use, the expected cost of the fix.
## Out of scope
`<sycl/sycl.hpp>` is not yet verified device-clean
(`syclbin_kernel_bundle.hpp` pulls `<filesystem>` → `<sstream>`; fixable
separately). `<sycl/ext/...>` cleanliness is partial: `bfloat16.hpp` and
the ESIMD type-traits headers are done, others untouched.
