Implement "measure read" function in pure assembly

[mmdriley]

Some of my rationale is in demos/asm/README.md, but the highlights are:

• MSVC does weird stuff with our intrinsics-based implementation and it makes me nervous
• The barriers necessary are so specific to the platform and particular context that it wasn't obviously great to have e.g. LFence() that was implemented on aarch64 and ppc64le.
• I want to make sure no compiler or flags update can change the latency value we observe on a given machine

This PR adds the assembly implementations with accompanying README, wires them up to be built, and rewrites callers to use them. There are also a couple of cleanup commits bundled in.

Some functions in instr.cc were left orphaned after this, so I removed them.

[mmdriley]

@asteinha this is ready for a look. Apologies for the large diff; here's one review strategy:

First, use "changes from" to review the cleanups separately. That's only ~15 lines but they might distract from the main review.

Then, in the main commit, a useful order might be:

• demos/asm/README.md and measurereadlatency.h
• The .S files for aarch64, ppc64le, and x86_64
• The .S file for x86, which is just a 32-bit-ification of the x86_64 version
• The .asm files, which are just copies of the corresponding .S files with most comments removed and one or two ABI tweaks.
• CMakeLists.txt changes to build the assembly
• Finally, source changes to use MeasureReadLatency

If the review starts feeling like a slog, let me know and I can split it up into different PRs. For example, I can bring things in one architecture at a time, then put in one last PR to build it and swap out the implementation.

Thank you for the review!

[mmdriley]

Rebased on #98 so we can make sure tests pass.