Wire Gemmini linalg lowering in IREE and add matmul IR dump

[101sparsh]

Summary

First step towards wiring Buddy’s Gemmini dialect into IREE and confirming that linalg.matmul can lower to a Gemmini op. This PR is mainly to get the wiring + IR dump in place so we have a concrete starting point for the full pipeline + backend work.

Changes
• Vendor Buddy Gemmini bits under compiler/src/iree/compiler/ThirdParty/buddy_gemmini/:
• Gemmini dialect (Gemmini.td, dialect/ops headers, registration)
• Lowering passes:
• LowerLinalgToGemmini (convert-linalg-to-gemmini)
• LowerGemmini skeleton (stub for later LLVM/legalization work)
• Simple test drivers (test-gemmini-opt.cpp, test-gemmini-simple.cpp)
• CMake wiring for the above.
• Register Gemmini and its passes in the IREE compiler so they are available through iree-opt.
• Add a tiny linalg matmul example under compiler/src/iree/compiler/Codegen/Common/test/gemmini_ir_dump.mlir.
• Add BUDDY_GEMMINI_INTEGRATION.md with notes on how this integration is structured and how to reproduce the IR dump.

How I tested

From the IREE build tree:

cat <<'EOF' >/tmp/test_matmul.mlir module { func.func @matmul(%a: tensor<4x4xf32>, %b: tensor<4x4xf32>) -> tensor<4x4xf32> { %c0 = arith.constant 0.0 : f32 %init = tensor.empty() : tensor<4x4xf32> %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4x4xf32>) -> tensor<4x4xf32> %0 = linalg.matmul ins(%a, %b : tensor<4x4xf32>, tensor<4x4xf32>) outs(%fill : tensor<4x4xf32>) -> tensor<4x4xf32> return %0 : tensor<4x4xf32> } } EOF

./build/tools/iree-opt /tmp/test_matmul.mlir \
--pass-pipeline='builtin.module(one-shot-bufferize,convert-linalg-to-gemmini)'
-o - | tee /tmp/gemmini_lowered.mlir | grep -n 'gemmini\.'

This produces:
11: gemmini.tile_matmul %1 %0 %alloc %alloc_0 : memref<4x4xf32, ...> memref<4x4xf32, ...> memref<4x4xf32> memref<4x4xi32>

and the full IR in /tmp/gemmini_lowered.mlir shows bufferization.to_buffer + allocs + a gemmini.tile_matmul op, confirming that convert-linalg-to-gemmini is wired and firing.

Next steps / scope
• This PR is only about wiring the dialect and proving we can lower a simple matmul to gemmini.tile_matmul via iree-opt.
• Not yet hooking Gemmini passes into the full IREE compile pipeline or any backend.
• Follow-up work:
• Decide the right place in the full pipeline (most likely early in the global opt stage) to insert convert-linalg-to-gemmini (and lower-gemmini once fully implemented).
• Make sure the full pipeline remains valid when these passes are enabled.
• Link Gemmini into the backend and start experimenting with real kernels / configs.