Add arm64 builder

[rengolin]

Adding a new ARM64 builder brings non-trivial changes:

• AWS Graviton instance needs to start/stop to avoid costs of running idle
• check_llvm.sh needs to know about both x86 and arm64 LLVM versions
• Integration tests checking for floating point values obviously fail
• Graviton node that is cheap to build TPP-MLIR isn't good to build LLVM
• We may want to separate LLVM builds from TPP-MLIR and keep the install dir zip in a cloud storage

(add more stuff here)

We may need to simplify execution, for example:

• only run clang release mode on Arm
• use the same builder for build and benchmark jobs
• have only benchmark jobs, since that builds release clang and test anyway

see https://github.com/plaidml/tpp-mlir/tree/arm64

[rengolin]

https://github.com/plaidml/tpp-mlir/actions/runs/12691058703/job/35373272875

Failed Tests (28):
  TPP_OPT :: BF16/Integration/matmul-pbf16.mlir
  TPP_OPT :: BF16/Integration/mlir-gen-bf16.mlir
  TPP_OPT :: BF16/Integration/mlp-all-bf16-tpprun.mlir
  TPP_OPT :: BF16/Integration/tpp-run-splat-shape.mlir
  TPP_OPT :: BF16/Integration/vnni-packing-chain.mlir
  TPP_OPT :: BF16/Integration/xsmm-brgemm-bf16.mlir
  TPP_OPT :: BF16/Integration/xsmm-gemm-bf16.mlir
  TPP_OPT :: BF16/Integration/xsmm-quarternary-bf16.mlir
  TPP_OPT :: BF16/Integration/xsmm-ternary-bf16.mlir
  TPP_OPT :: BF16/brgemm-tpp.mlir
  TPP_OPT :: BF16/brgemm-vnni.mlir
  TPP_OPT :: BF16/matmul-tiled-vnni.mlir
  TPP_OPT :: BF16/matmul-untiled-vnni.mlir
  TPP_OPT :: BF16/matmul-vnni.mlir
  TPP_OPT :: Conversion/LinalgToXsmm/linalg-to-brgemm.mlir
  TPP_OPT :: Conversion/LinalgToXsmm/linalg-to-gemm.mlir
  TPP_OPT :: Conversion/LinalgToXsmm/linalg-to-unary.mlir
  TPP_OPT :: Conversion/VectorToXsmm/vector-to-transpose.mlir
  TPP_OPT :: Conversion/XsmmToFunc/xsmm-to-func.mlir
  TPP_OPT :: Dialect/Xsmm/xsmm-dispatch-invoke.mlir
  TPP_OPT :: Integration/hoist-vector-transfer-brgemm.mlir
  TPP_OPT :: Integration/lower-pack-unpack-without-transpose.mlir
  TPP_OPT :: Integration/transpose-bf16.mlir
  TPP_OPT :: Integration/vector-contract-to-fma.mlir
  TPP_OPT :: Passes/DefaultPipeline/linalg.mlir
  TPP_OPT :: Passes/DefaultPipeline/vnni.mlir
  TPP_OPT :: Passes/DefaultPipeline/xsmm.mlir
  TPP_OPT :: Passes/xsmm-combine.mlir

Some errors I found:

• vector/fma tests fail diff by a large delta. Could be some tiling assumption that doesn't hold on Arm
• pass tests on VNNI fail verification as IR shape doesn't hold (error: 'xsmm.brgemm' op expect VNNI layout for operand B or invalid VNNI_B flags)
• pack/unpack integration test fails fpcmp with text instead of floats
• linalg-to-xsmm conversion tests don't lower VNNI operations
• all BF16 tests fail with no xsmm calls found or wrong VNNI layout
• BF16 MLP example fails with error: 'tensor.pack' op invalid tile factor or output size provided. Only full tiles are supported when padding_value is not set
• BF16 GEMM example failes with error: 'xsmm.gemm' op expect VNNI layout for operand: 2

[adam-smnk]

Interesting failures considering benchmarks run fine. It could indicate some tests are iffy(?)

[adam-smnk]

Looks like the main culprit is different VNNI blocking factor.
We usually run with factor 2. However, ARM uses 4 which is not compatible with (randomly) picked test sizes, hardcoded LIT checks, and/or prepacked arguments.