Summary

Wraps the existing layers/attention/scaled_dot_product_attention.go SDPA into a graph.Node[T] (FusedSDPA) so consumers (Wolf cross-attention, T3.1b) can compose it via graph.Builder without duplicating math.

• New layers/attention/fused_sdpa_node.go: FusedSDPA[T tensor.Numeric] implementing graph.Node[T]. OpType=\"FusedSDPA\", attributes head_dim + causal, no parameters, cached OutputShape().
• Forward accepts (Q,K,V) or (Q,K,V,mask); delegates to inner SDPA.
• Backward delegates to inner SDPA and appends a nil grad slot for the mask input when present, so input/grad indexing stays aligned.
• Options: WithFusedSDPABidirectional, WithFusedSDPAHeadCounts. Causal default mirrors existing SDPA convention (causal-on unless bidirectional).
• New layers/attention/fused_sdpa_node_test.go covers fp32/fp64 x {causal, bidirectional, masked} forward+backward equivalence vs the unfused ScaledDotProductAttention chain. Tolerances 1e-6 fp32 fwd / 1e-12 fp64 fwd / 1e-5 fp32 bwd / 1e-10 fp64 bwd. Stdlib testing only.

Downstream: T3.1b (Wolf swap) consumes this node.

Test plan

• go build ./...
• go vet ./...
• go test ./layers/attention/... -race -count=1 (1.29s, all green incl. new FusedSDPA cases)