ci: updates CI versions

[grzuy]

• Updates elixir/otp pairs
• DRY yml, collapse 3 jobs into one with matrix leverage
• more granular cache key

Pending:

• Fix OTP 26 stack_columns issue

[grzuy]

Test failure in OTP26 job seems to be legit failure in test/axon/compiler_test.exs:5388, I can reproduce also running it locally.

[grzuy]

Test failure in OTP26 job seems to be legit failure in test/axon/compiler_test.exs:5388, I can reproduce also running it locally.

And the EXLA job failure is consistent to the one already present in main.

[grzuy]

Test failure in OTP26 job seems to be legit failure in test/axon/compiler_test.exs:5388, I can reproduce also running it locally.

Seems that undeterministic map key order in OTP26 is affecting this test case

axon/test/axon/compiler_test.exs

Lines 5384 to 5391 in 713e440

	model = Axon.input("lazy_container") |> Axon.stack_columns(ignore: [:b])
	input = %LazyOnly{a: [[1]], b: [[2]], c: [[3]]}
	assert_equal(
	Axon.predict(model, %{}, %{"lazy_container" => input}),
	Nx.tensor([[1.0, 3.0]])
	)

which relies on Map.from_struct keeping order of struct fields...

axon/lib/axon/layers.ex

Lines 2526 to 2529 in 713e440

	deftransformp stack_columns_transform(container, ignore) do
	container
	|> Map.from_struct()
	|> Enum.reject(fn {k, _} -> k in ignore end)

Still not sure what should be the expected behavior of Axon.stack_columns for OPT26+... 🤔

[grzuy]

Still not sure what should be the expected behavior of Axon.stack_columns for OPT26+... 🤔

Hacky solution to keep Axon.stack_columns relying on the order of the container struct defstruct key list order? 3e8c5d8

[seanmor5]

Thanks!