Skip attention dropout in GATConv/GATv2Conv when dropout == 0

[Parvm1102]

NNlib.dropout(A, p) fetches the RNG (CUDA.default_rng() for CuArray) before it checks p, so dropout = 0 still pulls the CUDA RNG into the traced program, which Mooncake cannot differentiate. Every GATConv/GATv2Conv gradient on the GPU failed under Mooncake (#702).

MWE:

using CUDA, Flux, Mooncake, NNlib
x = CUDA.rand(Float32, 3, 4)
Flux.gradient(x -> sum(NNlib.dropout(x, 0.0)), Flux.AutoMooncake(), x)

TypeError: in typeassert, expected CoDual{IdDict{Any,Any}, IdDict{Any,Any}},
           got CoDual{IdDict{Any,Any}, NoFData}

Fix: skip the dropout call when p == 0 in gat_conv/gatv2_conv (also saves an RNG launch at inference). With it, GATConv and GATv2Conv pass all 96 GPU gradient sites under Mooncake (≤ 2.5e-7 vs Zygote); CPU tests unchanged. dropout > 0 still needs an upstream Mooncake rule for CUDA.default_rng(), tracked in #702.

[Parvm1102]

@CarloLucibello please review this.
Also my GSoC dashboard shows the mentor evaluation is still pending. Please submit it, the deadline is soon.