Crescent Island (cri) support for the XPU/SYCL build

[jiqing-feng]

Adds Intel Crescent Island (cri) as an AOT target for SYCL kernels.

Why it isn't just one more device in the list

cri needs a newer toolchain: ocloc 25.35 doesn't know the device, and
sycl-tla 0.9.1 has no cri kernels. Both are bumped (ocloc → 26.22.38646.4,
sycl-tla → 0.9.2), and AOT for cri is gated on DPC++ 2026.0.

The bigger problem is that a kernel compiled with __SYCL_TARGET_INTEL_GPU_CRI__
contains Xe35-only MXFP code that ocloc refuses to lower for pvc or bmg. So
such a translation unit can't share one AOT image across devices. The build now
detects that flag and picks the layout:

• without it — a single spir64_gen image serves pvc,bmg-g21,cri, which keeps
  the number of device compilations per TU down;
• with it — the generic image is restricted to cri, and pvc/bmg get separate
  intel_gpu_* alias targets.

Either way the .so also carries a spir64 JIT fallback, so a GPU that isn't
in the AOT list still runs.

Why the flag handling was rewritten

The SYCL flags used to be two hardcoded strings that dependencies patched with
string(REPLACE ...). That stops working once the target list is dynamic, and
it hid three failure modes that all produce a binary that builds fine and then
misbehaves:

• backend options must be repeated per target, and CMake deduplicates the
  identical -options strings, silently dropping every target but the first —
  SHELL: prevents this;
• -spirv-ext only reaches the image whose triple matches the exact
  -fsycl-targets alias, so it has to be emitted once per target;
• the SPIR-V translator replaces the extension list instead of merging, so the
  driver defaults are now probed with icpx -### and sycl-tla's extras appended.
  A failed probe is a hard error rather than a silent fallback.

Flags are now built from a few variables through a xpu_compose_sycl_flags()
macro, so a dependency changes a variable and recomposes.

Validation

Built flash-attn2 and megablocks from kernels-community against this
branch (torch213-cxx11-xpu20260-x86_64-linux) and checked numerics on both
BMG hardware and the CRI TBX simulator. Each case runs only the kernel on XPU;
data generation, the fp32 reference, MXFP4/MXFP8 dequantization and error
statistics stay on CPU.

• flash-attn2 — 8/8 pass (fix / varlen × fp16 / bf16 × causal / non-causal),
  max abs error 5.4e-04 (fp16) / 4.0e-03 (bf16).
• megablocks grouped GEMM — 5/5 pass (bf16 ±bias, fp16, mxfp4, mxfp8), max
  relative error 2.6e-03.