feat(tests): add 29 model tests for forward pass, LTI stability, generation, RMSNorm, causal mask, LoRA and loop-index embedding

[miheer-smk]

What this PR adds

Adds tests/test_model.py — 29 tests covering the core OpenMythos model.
The existing test suite only covered RoPE and the tokenizer.
This PR fills the gap by testing the model itself.

Tests added (9 categories)

• Model loading — GQA and MLA configs instantiate correctly
• Forward pass shape — logits are (B, T, vocab_size) for both attention types
• Batch sizes — parametrized over batch size 1, 2, 4
• LTI stability — spectral radius of A is in [0,1] at init and after extreme weight perturbation (core Parcae guarantee)
• Depth extrapolation — model works for n_loops=1,2,4,8 including beyond max_loop_iters
• Generation — output shape, prompt preservation, tokens within vocab
• RMSNorm — shape, unit RMS, learned weight scaling
• Causal mask — shape, upper triangle is -inf, lower triangle is 0
• Loop-index embedding — shape, differs per iteration, deterministic at t=0
• LoRA adapter — shape, valid across all loop indices

Related to #5

[Borda]

I believe you shall add minimal CI to run these and so prove they are correct... 🐰

[miheer-smk]

@Borda — done! Added .github/workflows/tests.yml which runs all 29 tests on Python 3.10 and 3.11 via GitHub Actions CI. Both passed successfully...

[miheer-smk]

@Borda @kyegomez — here's the CI proof. Both Python 3.10 and 3.11 pass all 29 tests on a clean machine:
✅ Python 3.10 — 29 passed
✅ Python 3.11 — 29 passed
Actions run: https://github.com/miheer-smk/OpenMythos/actions/runs/24804510337
Once merged, the workflow will run automatically on future PRs to the main repo too.