Share PrimitiveTypeToNbDLDataType with xla.

`PrimitiveTypeToNbDLDataType` is used in `LiteralToNdarray` in `xla_compiler` (jaxlib) which has python bindings for xla types. Create `xla_literal` which only contains bindings for `xla::Literal`.
- Moving `dlpack_support` out of jaxlib to third_party/tensorflow/compiler/xla/python
- Branched `dlpack` in jaxlib to third_party/tensorflow/compiler/xla/python so that the new xla dlpack target does not depend on pyclient.

PiperOrigin-RevId: 815912796

Author: hanrach9

tsl/platform/BUILD

@@ -27,6 +27,12 @@ load(
     "cc_library",
 )
 
+config_setting(
+    name = "libtpu_on_gce",
+    values = {"copt": "-DLIBTPU_ON_GCE"},
+    visibility = ["//visibility:public"],
+)
+
 package(
     # copybara:uncomment default_applicable_licenses = ["//tensorflow:license"],
     default_visibility = [

tsl/profiler/lib/BUILD

@@ -298,19 +298,13 @@ cc_library(
     visibility = ["//visibility:public"],
 )
 
-config_setting(
-    name = "libtpu_on_gce",
-    values = {"copt": "-DLIBTPU_ON_GCE"},
-    visibility = ["//visibility:public"],
-)
-
 cc_library(
     name = "nvtx_utils",
     hdrs = ["nvtx_utils.h"],
     visibility = ["//visibility:public"],
     deps = select({
         # Remove this once we clean up all the CUDA specific deps.
-        "libtpu_on_gce": [":nvtx_utils_external"],
+        "//tsl/platform:libtpu_on_gce": [":nvtx_utils_external"],
         # Place holder for internal dependencies
         "//conditions:default": [":nvtx_utils_impl"],
     }),