[Feat] GPU backends — burn-tch / burn-wgpu

[ax1s-x1zz]

Context

docs/ROADMAP.md Track D-1 (README Phase 6). Burn is backend-agnostic; Xazz currently compiles to the CPU backend only.

Scope

• burn-tch (CUDA) then burn-wgpu (cross-vendor) behind a feature flag / XAZZ_BACKEND
• Runtime detection with explicit CPU fallback warning

Acceptance

Same .xzz trains on CPU and CUDA with identical reported losses.

Dependencies

None (Burn API is backend-agnostic). Hardware availability gates timing.

---

See docs/ROADMAP.md for the full scale plan and execution order.