Old AVX sgemm code GPFs

[xloem]

I opened an issue for this, but the issue appears to have been deleted without comment somehow. I put a lot of work into formatting it to help anybody else engage it.

The default bolt testsuite is triggering a general protection fault in old code, apparently due to accessing memory from larger matrices with strides that are not a multiple of 32 bytes, at unaligned offsets.

I made a little progress on this, but an organised form of my change was only held in the lost issue.

I'll store a local copy of this issue, and update it to work it better.

[xloem]

Man, the world is not insane, just me. I opened the issue on the wrong repository: bolt/bolt#7967

I'll close this one and repost it here.