Define neural architectures in BASIC with packed tensors

[treeform]

GotA neural policies can define their architecture in BASIC using shared packed-tensor operations, so a new supported architecture no longer needs a custom native runner. Per-player weights and recurrent state stay inside the VM. BASIC scalar arithmetic remains deterministic integer/Q16.16; float32 is confined to tensors.

• Add checked, size-budgeted int32, Q16.16 and float32 tensor kernels, including dense, CSR, activations, elementwise math and gather/scatter.
• Implement all four current architectures in BASIC: Richard, David, Andre and Fly. Their tensor packages make no calls to the custom runner endpoints.
• Add ZIP conversion, synthetic runnable examples, API documentation and inference benchmarks. Existing named runners remain available.

Uses the merged Bassy #9, pinned to a02e6dd5889310b13d21725c891bde8d01539f08 in both dependency locks. Its contents match the locally tested dependency exactly.

Validation:

• Nim checks, full release suite with strict native BASIC compilation, lockstep regressions, converter tests and Bassy tests passed locally.
• Tensor tests passed with interpreted BASIC, native compilation and WASM. Tests cover recurrent resets, immutable weights, invalid handles/shapes, atomic failures, work and memory limits.
• Ten-player synthetic matches for all four architectures produced identical commands and simulation hashes against the named runners. WASM verified all 2,509 replay ticks per architecture; replay mismatches are still enforced.

Limits: private trained bot uploads were not tested. David's reference state differs by at most 5.96e-8, with at most one raw Q16.16 output unit of difference. Cross-platform float32 bitwise identity is not promised. Benchmarks are documented: small models pay extra host-call overhead, while the tested large David model was faster and Fly was about 1.55x slower than its specialized runner.