Tidy up old submodules and related scripts. (iree-org#3777)

* Remove `glslang` and `spirv_tools` from the Bazel `WORKSPACE` (both submodules were deleted)
* Delete old, outdated `populate_reference_repo.sh` script
* Remove [Marl](https://github.com/google/marl) submodule

.gitmodules

@@ -44,9 +44,6 @@
 [submodule "third_party/mlir-emitc"]
 	path = third_party/mlir-emitc
 	url = https://github.com/iml130/mlir-emitc.git
-[submodule "third_party/marl"]
-	path = third_party/marl
-	url = https://github.com/google/marl.git
 [submodule "third_party/flatcc"]
 	path = third_party/flatcc
 	url = https://github.com/dvidelabs/flatcc.git

SUBMODULE_VERSIONS

@@ -7,7 +7,6 @@ a5d9d0f7d368054fd1691aedf1db4116efcc233e third_party/flatbuffers
 f2fb48c3b3d79a75a88a99fba6576b25d42ec528 third_party/googletest
 e3f662922d8340a7fd46a4f75cad0e72163bf219 third_party/llvm-bazel
 f147f59cd377a6be68e5ca5c343eb11df8e7ee6f third_party/llvm-project
-17b12a4481daa150e2d1ea3ada086b551b856707 third_party/marl
 9e970e18057e80e8af07f96d73714e395bac98d1 third_party/mlir-emitc
 d8c7ee00a687ac369e62e2032514a93a9b413502 third_party/pybind11
 d79362c24fd70eab3196672273dbfd8f0dd6124c third_party/ruy

WORKSPACE

@@ -227,19 +227,6 @@ maybe(
     path = "third_party/vulkan_memory_allocator",
 )
 
-maybe(
-    new_local_repository,
-    name = "glslang",
-    build_file = "build_tools/third_party/glslang/BUILD.overlay",
-    path = "third_party/glslang",
-)
-
-maybe(
-    local_repository,
-    name = "spirv_tools",
-    path = "third_party/spirv_tools",
-)
-
 maybe(
     local_repository,
     name = "spirv_headers",

docs/design_roadmap.md

@@ -717,8 +717,8 @@ tile size is defined by a library call contract, heuristics, or empirically is
 TBD, but expect workgroup sizes in the thousands to millions of invocations vs.
 normal GPU workgroup sizes in the dozens to hundreds of invocations.
 
-To achieve this style of scheduling efficiently we'll likely use
-[marl](https://github.com/google/marl) as the scheduler. This provides
+To achieve this style of scheduling efficiently we'll likely use something like
+[marl](https://github.com/google/marl) as the scheduler. Marl provides
 cross-platform low-overhead fibers and is compatible with this style of
 scheduling as it was built for the Swiftshader software rasterizer.