PR tensorflow#14477: Allow to build multihost_hlo_runner for CPU backend

Imported from GitHub PR openxla/xla#14477

Currently, Bazel skips building `multihost_hlo_runner:hlo_runner_main` on CPU because the target is tagged with `tf_gpu_tests_tags`.

This tag should only be applied to the dedicated `hlo_runner_main_gpu` target.

The regular `hlo_runner_main` works fine on CPU (when `--xla_force_host_platform_device_count=N` is used)

@bchetioui @ddunl @beckerhe Can you have a look?
Copybara import of the project:

--
a0d2129bfe74f2f9133f9e1c6b04e06cea6d0e5d by Alexander Pivovarov <pivovaa@amazon.com>:

Allow to build multihost_hlo_runner for CPU backend

Merging this change closes tensorflow#14477

PiperOrigin-RevId: 650160788

third_party/xla/xla/tools/multihost_hlo_runner/BUILD

@@ -31,7 +31,7 @@ cc_library(
     compatible_with = None,
     tags = [
         "nomac",
-    ] + tf_gpu_tests_tags(),
+    ],
     deps = [
         ":create_client",
         ":functional_hlo_runner",
@@ -73,7 +73,7 @@ xla_cc_binary(
     testonly = True,
     tags = [
         "nomac",
-    ] + tf_gpu_tests_tags(),
+    ],
     deps = [
         ":hlo_runner_main_lib",
     ],