PGOLF-627: add Parameter Golf train-time ReLU^2 graph and runtime support

[AtlantisPleb]

Parent: #194, #183
Blocked by: none

Roadmap lane: Epic 6, Rust-only single-device trainer closure.

Why this exists:

• the current public crates/psionic-models/src/parameter_golf.rs baseline MLP uses exact ReLU^2 activation (max(x, 0)^2), not SiLU
• psionic-ir still does not expose an exact public train-time relu, clamp_min, maximum, or equivalent decomposition that can lower that baseline MLP honestly on the Rust-owned graph path
• without an exact ReLU^2 answer, #269 cannot finish the full baseline forward-logits graph and seeded backward path honestly

Scope:

• add one exact public train-time answer for the Parameter Golf ReLU^2 MLP activation on the bounded graph/autodiff/runtime path
• keep the first answer honest if it is bounded or host-orchestrated rather than fused
• add targeted parity tests against the existing Parameter Golf reference math or a direct reference implementation

Exit criteria:

• the public graph/autodiff/runtime stack can execute and differentiate one bounded exact ReLU^2 train-time path needed by the Parameter Golf baseline
• the landed surface is sufficient to remove ReLU^2 as an implicit blocker for #269 and #194
• docs or machine-readable truth are updated where the new primitive changes the honest trainer posture

[AtlantisPleb]

Completed in 787f4be on main.

Summary:

• added exact relu_squared and relu_squared_backward backend-extension ops plus stable labels and kinds in psionic-core
• registered the new ops in psionic-ir, including operator schemas, meta-execution, graph-builder entrypoints, dense reference evaluation, and reverse-mode ReLU^2 gradient lowering
• added bounded CPU reference execution and CUDA runtime execution for forward and seeded backward ReLU^2 graphs
• corrected the Parameter Golf roadmap so #270 no longer claims it retired the baseline MLP blocker; the actual baseline activation gap is now closed by #271

Validation run:

• cargo test -p psionic-ir relu_squared -- --nocapture
• cargo test -p psionic-backend-cuda relu_squared -- --nocapture
• cargo test -p psionic-backend-cpu backend_selection_ -- --nocapture

This retires the real baseline MLP activation blocker for the Rust-native Parameter Golf train graph without claiming logits, loss, or end-to-end trainer closure.