Refactor rms_norm transform script to follow mlir-air layernorm prototype

[erwei-xilinx]

Summary

Refactors examples/rms_norm/transform_aie2p.mlir to follow the structure of the mlir-air xrt prototype test/xrt/43_triton_layernorm/transform_aie2p.mlir exactly.

Fixes the spurious stderr reported during testing of #64 on Windows aie2p:
```
loc("-":83:11): error: application of transform.air.copy_to_dma expected to produce 1 results (actually produced 0).
```

Root cause

The previous script used `transform.air.linalg_promote` (post-bufferize) to stage L2 subviews into L1. When the linalg op's destination is itself an L2 subview, `linalg_promote` emits `memref.copy %sv, %sv` (literal SSA self-copy). `transform.air.copy_to_dma` then detects each self-copy in `CopyToDmaOp::applyToOne`, erases the op, and returns success without pushing a result handle — violating the framework's 1-result contract for `applyToOne` and producing the diagnostic above. Reproduces deterministically on Linux aie2p.

Fix

Replace the post-bufferize `linalg_promote` mechanism with the layernorm prototype's pre-bufferize approach:

• Per-op `bufferize_to_allocation {memory_space = 2}` for each linalg op (fills, intermediate generic, reduce, output generic) → L1 alloc as the bufferization destination
• `promote_tensor to 2` for the input tensor → L1 input copy
• Drop `linalg_promote` and the custom `fuse_multi_op_linalg` of sq+reduce
• Add `generalize` of remaining `linalg.reduce` (delayed to PHASE 9), `convert_divf_sqrt_to_rsqrt`, and `broadcast_before_unary` for math.rsqrt — all matching the layernorm prototype

Self-copies are never emitted, so the diagnostic disappears structurally rather than via a workaround. Resulting AIR shows clean separated loops (square → reduce → output) with a single shared L1 input alloc.

Test plan

• `scripts/run_tests.py --device aie2p` on local NPU2 (Strix): 17/18 pass (only pre-existing matvec unrelated)
• Standalone `python examples/rms_norm/rms_norm.py` on aie2p: pass, no stderr diagnostics
• Compilation produces `aie.elf` + `full_elf_config.json` cleanly
• Verify on Windows aie2p (the platform where Add Windows wheel release job to nightly-wheels CI #64 reported it) — expected to pass since fix is host-OS-independent
• aie2 path unaffected (no `transform_aie2.mlir` exists for this kernel; aie2p-only)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR refactors the AIE2P MLIR transform sequence for the rms_norm example to mirror the mlir-air layernorm XRT prototype structure, replacing the prior post-bufferize linalg_promote-based L1 staging approach that could generate SSA self-copies and trigger transform.air.copy_to_dma result-contract diagnostics.

Changes:

• Reworks the transform pipeline into phased steps (canonicalize/fuse, dataflow navigation, tiling+fusion, pre-bufferize allocations/promotions, bufferize, vectorization prep, herd/DMA/vectorization, type casts).
• Replaces post-bufferize transform.air.linalg_promote staging with pre-bufferize bufferize_to_allocation {memory_space = 2} and promote_tensor to 2 to avoid self-copy patterns.
• Adds prototype-aligned transforms: delayed generalize of remaining linalg.reduce, convert_divf_sqrt_to_rsqrt, and broadcast_before_unary for math.rsqrt.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.