feat(model-ports): add jamba_tiny_dev

[DarthCeltic]

Adds ai21labs/Jamba-tiny-dev (319M, Apache-2.0) as a new model-port claim, introducing the Jamba execution family (hybrid Mamba+Transformer MoE, distinct from other hybrid families already on the board).

Self-converted GGUF (no pre-made quant exists for this checkpoint), published as a GitHub Release since no HF upload token is available. Locally verified: arch=jamba loads cleanly (16 layers, both recurrent-state and KV caches allocated as expected for the hybrid architecture), clean 1009-node compute graph, real PPL = 18.4574 +/- 3.82 against WikiText-2 (within this campaign's normal range).

Full provenance, conversion steps, and an honest note on verification tier (CPU-backend, not full ET-SoC1 sysemu — see recipe for why) in ported_models/jamba_tiny_dev/docs/RECIPE.md.

[AFOliveira]

The recipe honestly reports CPU-only verification. To receive model-port credit, Jamba still needs a real full-offload ET-SoC1 execution, a committed deterministic oracle/expected output with an explicit comparison threshold, and immutable release provenance. Host loading and CPU PPL alone are insufficient. Maintainers will then review/adopt the Jamba identity/contract and runner policy.