Test upsample testsuite against multiple eltypes

[CarloLucibello]

Addresses the T = Float32 # TODO test against all supported eltypes for each backend item catalogued in #702.

The forward/gradient testsets in test/testsuite/upsample.jl (upsample_nearest, linear/bilinear/trilinear) are now run in a loop over the float eltypes each backend supports:

float_types = Symbol(Backend) == :MetalBackend ? [Float16, Float32] : [Float16, Float32, Float64]
@testset "eltype = $T" for T in float_types

This mirrors the per-backend eltype pattern already used in test/testsuite/scatter.jl. Metal drops Float64 since it has none natively.

Details

• Float16/Float64 forward + eltype-preservation are now covered (previously only Float32). isapprox automatically relaxes rtol to the less-precise operand, so the existing ≈ ground-truth comparisons pass for Float16 unchanged.
• Gradients are checked only at Float32/Float64. Finite-difference gradient checking is unreliable at half precision (observed 3–17% error), so for Float16 only the forward pass runs — via gradtest's skip=true (CPU) / gputest's checkgrad=false (GPU).
• GPU safety: the explicit ∇upsample_bilinear/∇upsample_trilinear correctness checks invoke the gradient kernels, which use @atomic adds (half-precision atomics aren't universally available on GPU), so they're gated behind the same condition. Float16 stays forward-only — no atomics — on every backend.
• The Rational/UInt8 forward tests moved out of the per-eltype loop into their own testset so they run once instead of once per float eltype. pixel_shuffle and the complex-valued tests are unchanged.

Testing

CPU testsuite is green: 129 pass, 5 intentional Float16 gradient skips, 0 failures. GPU backends were not run locally.

🤖 Generated with Claude Code