Extend lightweight herd cloning to tiled unroll path in air-opt-shim-dma-bds

[erwei-xilinx]

Context

PR #1535 introduces lightweight herd cloning during shim-level loop unrolling in loopUnrollFullWithAsyncTokenPreserved (mlir/lib/Util/Dependency.cpp). When unrolling loops that contain air.SegmentOp or air.HerdOp, it creates empty herd shells via OperationState and only clones channel ops + their transitive dependencies, skipping heavy compute ops (vector, arith, linalg). This avoids O(N × body_size) IR explosion.

Limitation

The lightweight unroller only fires when annotateFn is null — the non-tiled unroll path. The tiled path in AIROptimizeShimDMABDs (triggered by non-trivial shim-dma-tile-sizes like 2,2 or 4,4) uses loopUnrollByFactor with an annotateFn callback to tag unrolled iterations. This path still performs full deep-clone unrolling.

The annotateFn guard exists because loopUnrollByFactor is an upstream MLIR utility that doesn't support pluggable clone strategies. Extending lightweight cloning to this path requires either:

1. Wrapping loopUnrollByFactor to accept a custom clone callback, or
2. Performing a post-unroll strip (clone-then-strip) specifically for the tiled path — noting that a naive strip crashes due to cross-region use-after-free (see T002 postmortem in PR Lightweight herd cloning during shim DMA BD loop unrolling #1535 discussion)

Impact

For the default aircc invocation (tile sizes 1,1), this limitation has no effect — the shim BD pass doesn't tile or unroll. For workloads that use explicit tiling (e.g., flash attention with large trip counts and non-default tile sizes), the tiled unroll path still sees the full IR explosion.