Add GNN message-passing examples for the einsum JIT

[MesTTo]

This PR is stacked on #115.

Adds two examples that exercise the einsum-f32 JIT on graph neural-network message passing: gnn_stress (a single large layer, comparing the VM and JIT paths and checking they match) and gnn_million (a million-edge multi-layer sweep). Behind the jit feature.

They double as a sanity check on the JIT backend: gnn_stress asserts the JIT result matches the VM (max_abs_diff 0) and reports the speedup.

Validation:

• CARGO_HOME=/home/user/Dev/.cargo-isolated RUSTFLAGS='-C target-cpu=native -Awarnings' cargo +nightly build -p linalg --features jit --examples
• cargo +nightly run -p linalg --features jit --example gnn_stress (JIT matches VM, ~146x faster on this machine)
• git diff --check

[Adam-Vandervorst]

Hmm I don't see why this PR has to be stacked; can't all /linalg changes be applied independently? I'd love more benchmarks here.

[MesTTo]

the /linalg PRs are independent of each other... #111 (blas backend), #112 (einsum output-shape inference) and #115 (jit rework) each apply straight on main. this one only sits on #115 because the examples call the reworked api (EinsumF32Plan / einsum_auto), which doesn't exist before it, so once #115 lands this drops to just the two example files.

if you'd rather, i can retarget the examples at the current jit api so they stand alone, or fold them into #115... whichever reads better for you.

on more benchmarks: i have a crossover bench (openblas vs the jit as density increases) that i left out only because it needs both #111 and #115. i can add it the moment those are in, or bundle it now if you want it sooner.

[MesTTo]

folded these into #115 so the jit rework ships self-contained with its examples and benchmarks, rather than as a separate stacked pr. closing this one... the gnn examples + graph/perf benches now live in #115 and apply independently on main. thanks for the nudge.