reland: [Apple] Support benchmarking multiple Mach-O slices (#402)

This is enabled through a new CMake variable `TEST_SUITE_RUN_ARCHS` that allows
users to specify which subset arches from CMAKE_OSX_ARCHITECTURES they'd like to
run the tests for. By default, when there's two or more, we emit
foo.<slice>.test files for all of them. Existing names are preserved when
TEST_SUITE_RUN_ARCHS is a list of size zero or one.

Implementation-wise, we extend the `timeit` binary to support `-arch <slice>`,
much like macOS's `arch` utility, and plumb in the requested slice name via a
new `ARCH:` directive in the .test files.

This is a re-land of #398, which was reverted in f0e4ef0. Now includes a couple of fixes for old SDKs.