Minimal fix of #1556, remove eltype checks

[mcabbott]

Fix #1556, closes #1557.

Tests of this conversion behaviour are left skipped for now, that seems the minimal step. Perhaps friendly warnings if you're about to accidentally promote to Float64 can be combined with not rejecting tracking types. But that need not be designed in a rush.

The complete solution of #1556 should include tests of gradients through the re of destructure. At the moment it is (I think) only tested alongside loadparams! as an "offline" utility, but clearly it's also useful to map a vector of parameters to a model inside a solver, etc. Perhaps it can be made faster, too.

[darsnack]

Okay well both this PR and #1557 are "approved" by me, but we can only merge one. Personally, there are aspects of both PRs that I would mix and match. #1557 removes the inner constructor for Dense and widens the weight type, which I think is right. But this PR doesn't make changes to the the main user-facing constructor for Dense while #1557 does (and I think those changes increase the surface area for errors). Other than that, both PRs fix #1556.

[DhairyaLGandhi]

Thanks for the PR! Appreciate the quick response, and agree that we don't want to rush.

This solves the particular MWE of #1556, on balance, we also add some other cleanups in the other PR, ie loosening the type restriction and removing the inner constructor which we want. These are connected issues and both good changes. What I am inclined to do is merge both PRs, since #1557 is effectively a superset.

bors r+

[DhairyaLGandhi]

bors ping

[bors]

pong

[darsnack]

Probably the test @mcabbott mentioned can be added to the "downstream CI" @DhairyaLGandhi is adding in #1555. We should also add some tests for "non-conventional" parameter array types.

[mcabbott]

and agree that we don't want to rush.

Then why did you?

Arguments for changes other than fixing #1556 really need to stand on their own merits. They were extensively discussed at the time. I think you should revert #1557, and (if you want to discuss other changes) open a normal discussion.

[darsnack]

Never mind, I misunderstood the downstream CI. We should add the tests mentioned here directly to Flux.