Function fuse_linear 

pub fn fuse_linear(graph: &SharedGraph, roots: &[NodeId]) -> usize

Fuse MatMul → Add(rank-1 bias) patterns into Op::FusedLinear (Phase 6d Track 3). Walks the graph; for each Add whose LHS is a MatMul and whose RHS is a rank-1 bias whose length equals the matmul output’s last dim, emits a fresh FusedLinear node and remaps consumers of the Add to it.

Conservative: only fires when the Add is the sole consumer of the MatMul. Otherwise we’d be creating a duplicate matmul computation. CSE doesn’t help here because MatMul and MatMul-inside-FusedLinear aren’t structurally equal at the IR level — backends with truly fused kernels are the ones that benefit, so we’d rather skip the fusion than waste the work.

Returns the count of fusions applied.