README.Add Status Macros

add-status-macros.patch adds ASSERT_OK, EXPECT_OK, ASSERT_OK_AND_ASSIGN to gmock.h so that the header's provided functionality matches internal gmock.

What other things have we tried?

1. Introducing a custom header to be used in OSS instead of gmock/gmock.h.

   The export-to-OSS process imposes a few restrictions. Notably, header rewrite has to be reversible, so we need a 1:1 mapping between headers used internally and in OSS.

   If we introduced a custom header to be used in OSS instead of gmock, it would have to take the place of current rewrite of internal gmock to gmock/gmock.h. This means, any use of gmock/gmock.h in OSS XLA code can no longer map to internal gmock. We'd have to ban the header.

   Therefore, updating OSS gmock/gmock.h seems necessary.

2. Patching in the extra macros to gmock/gmock.h by including absl/status/status_macros.h.

   This introduces a circular dependency between absl and gmock which makes bazel strongly opposed to the idea.

3. Introducing a googletest bazel module wrapper.

   This would be a module would proxy all gmock/gmock.h within XLA without additional patching of googletest. However, having multiple sources of the same gmock/gmock.h header path only works sometimes. The order of include paths emitted by bazel depends on the target definition and ordering of dependencies, so it ends up working in some case and not in others.

4. Expanding 3. by renaming googletest's gmock.h to gmock.upstream.h to avoid header name conflicts.

   gmock/gmock.h is also included by googletest itself, so redirecting it to gmock/gmock.upstream.h is needed. That boils down to even more brittle patching.

Overall, the add-status-macros.patch change is the least invasive one that works.