Test all buildkite runs on LTS and latest

[kshyatt]

No description provided.

[christiangnrd]

Lgtm

[christiangnrd]

Bump version?

[maleadt]

Bump version?

Bit weird to do this as part of this PR, but OK.

---

I was going to suggest trimming the test suite a bit, since this is increasing CI load, but then noticed that macOS runs finish in 10 minutes. Really surprising that e.g. the gpuarrays/reductions/sum prod test suite is 3 times faster on the macOS workers (M1s) than it is on gpuci (AMD EPYC 7402 CPU). The M1 does have 1.7x higher single-threaded performance according to PassMark, which surprised me, but that's still a far cry from 3x.

[christiangnrd]

Bit weird to do this as part of this PR, but OK.

Noted for future reviews from me.

Regarding the M1 vs Epyc, Metal supports fewer types so it runs fewer tests. I did some math for my local machines, and for gpuarrays/reductions/sum prod from your example, the tests per second ratio between metal and cuda is the same as the ratio between the passmark scores for the respective CPUs. I guess even though they're running on GPU the tests are so small that it's all CPU-limited regardless

[luraess]

yay all CI passes now