Add MIOpen (AMDGPU) fast path for batchnorm

[CarloLucibello]

Following the addition of the functional normalization operators (batchnorm, instancenorm, groupnorm, layernorm) in NNlib v0.9.41, the cuDNN-accelerated batchnorm lives in NNlibCUDACUDNNExt and is selected automatically for CuArrays. There is no equivalent MIOpen fast path for AMDGPU ROCArrays — they fall back to the generic implementation.

Flux is migrating its normalization layers to wrap NNlib.batchnorm (FluxML/Flux.jl#2700) and, as part of that, is removing its own AMDGPU MIOpen BatchNorm specialization (previously in ext/FluxAMDGPUExt/batchnorm.jl). To keep the accelerated AMD path available, it would be good to move that logic into NNlib as an AMDGPU/MIOpen extension, mirroring NNlibCUDACUDNNExt.

The code removed from Flux, as a starting point:

function _amdgpu_batchnorm(x, γ, β; μ, σ², ϵ, within_grad::Bool)
    if within_grad
        return AMDGPU.MIOpen.batchnorm_training(x, γ, β, μ, σ²; ϵ=Float64(ϵ), iteration=0) # TODO iteration
    else
        return AMDGPU.MIOpen.batchnorm_inference(x, γ, β, μ, σ²; ϵ=Float64(ϵ))
    end
end

function ChainRulesCore.rrule(::typeof(_amdgpu_batchnorm), x, γ, β; μ, σ², ϵ, within_grad::Bool)
    y, μ_saved, ν_saved = _amdgpu_batchnorm(x, γ, β; μ, σ², ϵ, within_grad)
    function _batchnorm_pullback(Δ)
        dx, dγ, dβ = AMDGPU.MIOpen.∇batchnorm(unthunk(Δ), x, γ, β, μ_saved, ν_saved)
        (NoTangent(), dx, dγ, dβ)
    end
    y, _batchnorm_pullback
end

Ideally the NNlib method would implement the full batchnorm(g, b, x, running_mean, running_var, momentum; eps, training, track_stats) signature (matching the generic and cuDNN methods) and its ∇batchnorm, honoring training / track_stats / running-statistic updates rather than relying solely on within_gradient as the old Flux code did. It should also cover the 2D/4D/5D input shapes and address the iteration TODO.