Add stream-dse fused SwiGLU-prefill operator

[asyms]

Adds SwiGLUPrefillStream, a fused SwiGLU-prefill operator whose single MLIR design (gate/up GEMMs + SiLU + elementwise-mul + down GEMM) is generated by stream-dse and compiled to one xclbin, instead of chaining separately-compiled sub-operators.

Its per-kernel operand layouts (the tiled-strided DMA tiling) are authored on the IRON side and injected into stream-dse code generation via optimize_allocation_co(kernels=...) — so IRON owns the layouts while stream keeps kernel construction and the MLIR rewrite, instead of the layouts being hand-copied on both sides.

Added

• SwiGLUPrefillStream (iron/operators/swiglu_prefill_stream/): fused stream-dse design → one xclbin; MLIR generated at build time by stream_design.py.
• iron.common.layout: a TiledStridedLayout type (with to_snaxc()) for handing IRON-authored operand layouts to stream-dse.
• stream_kernels.py: injects IRON's operand layouts into codegen through the kernels= override, replacing only operand_layouts() on stream's own kernels (requires stream-dse ≥ 1.13.4).
• requirements_stream.txt (optional dependency stream-dse>=1.13.4); the operator's test skips when stream-dse is absent.
• Minimal demo under demos/swiglu_prefill_stream/.

Changed

• Importing iron.operators no longer requires an NPU runtime: lazy XRT/pyxrt import and PEP 562 lazy operator exports, so the package loads (and tests collect) on hosts without XRT/pyxrt.

Removed

• None.

Running the demo

Prerequisites: the XDNA driver + XRT installed (/opt/xilinx/xrt) and an npu2 device. From a fresh clone of this branch:

python3 -m venv .venv && source .venv/bin/activate
source /opt/xilinx/xrt/setup.sh            # provides pyxrt
pip install --upgrade pip
pip install -r requirements.txt            # IRON + mlir_aie/llvm-aie toolchain + torch
pip install -r requirements_stream.txt     # stream-dse>=1.13.4 (PyPI)
stream-setup-aie                           # required: installs snaxc / xdsl-aie / aie-python-extras
python demos/swiglu_prefill_stream/demo.py

This generates the fused design with stream-dse, compiles it to an xclbin, and runs it once on the NPU (≈2 ms for the 256×512×2048 shape). stream-setup-aie is required: it installs the AIE codegen packages stream-dse needs that cannot be plain PyPI dependencies.

Licensing note

The new IRON-side files — iron/common/layout.py, iron/operators/swiglu_prefill_stream/stream_kernels.py, and demos/swiglu_prefill_stream/demo.py — carry a KU Leuven (MICAS) copyright header (Apache-2.0), as they were authored by MICAS; all other touched files keep their existing AMD headers. We can discuss this further.

PR Merge Checklist

1. The PR is rebased on the latest devel commit and pointing to devel.
2. Your PR has been reviewed and approved.
3. All checks are passing.

[andrej]

Hi Arne, sorry for the CI failures, if you rebase on #125 maybe once it's merged hopefully these should pass