Julia 1.14 compatibility; CI step selection

[maleadt]

On main, nightly (1.14) fails in gpuarrays/uniformscaling (jl_genericmemory_copyto in device code) and more.

• CI: [only X]/[skip X] take comma-separated step tags, as in GPUCompiler: julia, nightly (was pre), special, validation, docs, benchmarks, and tests for all test steps. Draft PRs only test the newest release (1.13) and nightly. Older releases and the special steps run once the PR is ready. Steps share their definition through YAML anchors.
• Consistent overlays: plain overlays taint every call that can reach them, including through error paths, which blocks concrete evaluation. Since 1.14 guards array growth with checked_add (base: Guard allocation arithmetic from overflow JuliaLang/julia#62797), that broke keyword-argument handling in kernels, e.g. range(start; step, length) in LinearAlgebra's diagind (A + I). @device_override now uses @consistent_overlay, as CUDA.jl and AMDGPU.jl do. The Random.Sampler override returns a different sampler than the host method, so it stays a plain overlay; otherwise concrete evaluation substitutes the host's SamplerRangeNDL.
• RNG: @invoke randn(rng::AbstractRNG, T::Type{<:AbstractFloat}) built a UnionAll at run time on 1.14 (type system: Add TypeEgal kind JuliaLang/julia#62001). The signature is now a constant.
• Type broadcasts: 1.14 specializes type arguments on Core.TypeEgal{T}, which the Broadcasted{…, Type{T}} adaptor rule didn't match, so Float32.(x) failed to compile. It now matches <:Type{T}.
• Functions capturing a type: LinearAlgebra's triangular det now reduces with Base.Fix1(convert, T) (Check that matrix is square in det triangular fast-path JuliaLang/LinearAlgebra.jl#1658). Fix1/Fix2 capturing a type are converted to closures.
• Argument encoding bug (latent, exposed by the previous item): the unconverted argument type decided whether an argument takes a slot. A Fix1 that converts to a ghost closure shifted every later argument, so sum(Base.Fix1(convert, Float32), x) returned 0. The slot is now decided on the converted type.

The full test suite passes locally on nightly (1.14.0-DEV.3363, M3 Pro), with JuliaGPU/GPUCompiler.jl#956 for the write-barrier fix. That's in GPUCompiler 2.9, which this PR now requires.

[codecov]

Codecov Report

❌ Patch coverage is 78.57143% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 85.04%. Comparing base (26c9e7d) to head (c5e5377).
⚠️ Report is 1 commits behind head on main.

Files with missing lines	Patch %	Lines
src/compiler/execution.jl	72.72%	3 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #979      +/-   ##
==========================================
- Coverage   85.37%   85.04%   -0.33%     
==========================================
  Files          77       89      +12     
  Lines        5660     6387     +727     
==========================================
+ Hits         4832     5432     +600     
- Misses        828      955     +127     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.