CI fails in the build-cuda step

[patowen]

The CI process currently fails every PR in the build-cuda step due to the changes in #1112. I believe there are multiple ways to fix this:

• Use the latest nightly version instead of a specific version when building for CUDA
• Update the Rust nightly version (much like what was done in 229c137)
• Use the latest Rust stable version instead of nightly (I haven't researched if this is possible, and I expect it's not)
• Revert Make "Point::new" a const fn #1112 (not the ideal solution)
• Add a non-const version specifically for targeting CUDA, probably using the cfg attribute (as suggested in [Sparse] Add clear_triplets() #1141 (comment))

I believe the code to modify is at

nalgebra/.github/workflows/nalgebra-ci-build.yml

Lines 129 to 133 in 7adecde

	- name: Install nightly-2021-12-04
	uses: actions-rs/toolchain@v1
	with:
	toolchain: nightly-2021-12-04
	override: true

I would create a PR, but I don't have enough context to know which bullet point is preferred.

[sebcrozet]

Hi! I’ll take care of this in #1156. We can’t use stable or update the nigthly version because rust-cuda requires a specific version of the compiler. The clean fix would be to revert Point::new being const, but it feels unfortunate to remove this for everybody where it’s only really an issue when using cuda.

So let’s just add a non-const version behind a cfg feature flag. This does go against the additive nature of cargo features, but I think it’s fine for now.

[patowen]

Oh, got it. This limitation is clearly written at https://github.com/Rust-GPU/Rust-CUDA/blob/master/guide/src/guide/getting_started.md#rust-toolchain. Thanks for the fix!