gemm: GPU acceptance tests vs cuBLAS + finite-diff

Forward: hand-written tiled GEMM vs cuBLAS sgemm on random matrices
(square / non-tile-aligned rect / 256³), max relative error < 1e-3, using
the row-major⟺col-major identity to drive cuBLAS without explicit
transposes. Backward: scalar loss L = sum(W∘C) (so dC = W), dA/dB from
matmul_backward checked against the finite-diff harness. Gated behind
not(no_cuda).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

crates/xtrain-tensor/Cargo.toml

@@ -7,3 +7,6 @@ edition.workspace = true
 xtrain-cuda = { path = "../xtrain-cuda" }
 half.workspace = true
 smallvec.workspace = true
+
+[dev-dependencies]
+xtrain-autodiff = { path = "../xtrain-autodiff" }