Remove attention transform dialect e2e tests (iree-org#17682)

With iree-org#17681 , the state of
attention codegen for llvm-cpu is at a good enough place where we don't
need to rely on a transform dialect spec (at least for llvm-cpu)
anymore.

This patch removes e2e tests for the attention transform dialect spec as
it adds more burden on a path that will probably not be maintained in
the future. There are some e2e correctness tests in e2e/linalg_ext/ that
check correctness for small test cases and pkgci running sdxl +
attention tests.

tests/transform_dialect/cpu/BUILD.bazel

@@ -14,7 +14,6 @@ package(
 iree_lit_test_suite(
     name = "lit",
     srcs = [
-        "attention.mlir",
         "contraction-packing.mlir",
         "contraction-packing-and-dispatch.mlir",
         # DISABLED: incorrectly assuming default flag values.
@@ -28,7 +27,6 @@ iree_lit_test_suite(
     # transform dialect spec files are MLIR files that specify a transformation,
     # they need to be included as data.
     data = [
-        "attention_codegen_spec.mlir",
         "matmul_codegen_default_spec.mlir",
         "transform_library.mlir",
     ],

tests/transform_dialect/cpu/CMakeLists.txt

@@ -14,7 +14,6 @@ iree_lit_test_suite(
   NAME
     lit
   SRCS
-    "attention.mlir"
     "contraction-packing-and-dispatch.mlir"
     "contraction-packing.mlir"
     "fold_tensor_slice_into_transfer.mlir"
@@ -27,7 +26,6 @@ iree_lit_test_suite(
     iree-opt
     iree-run-module
   DATA
-    attention_codegen_spec.mlir
     matmul_codegen_default_spec.mlir
     transform_library.mlir
   LABELS