Add Evo2-7B contrib model (StripedHyena 2 DNA language model)

[jimburtoft]

Summary

• Adds Arc Institute's Evo 2 (7B, StripedHyena 2 architecture) DNA language model to contrib
• Hybrid SSM+Attention architecture with custom NeuronFFT (Stockham radix-2) replacing unsupported torch.fft.*
• Vanilla torch_neuronx.trace() compilation with block-by-block prefill (32 NEFFs) and batched decode with HBM-resident state

Model Details

• Source: arcinstitute/evo2 (7B params, Apache 2.0)
• Architecture: 32-layer StripedHyena 2 (9 HCS + 9 HCM + 9 HCL/SSM + 5 ATT blocks)
• Tokenizer: Single-nucleotide (A, C, G, T), vocab=512
• Instance: trn2.3xlarge (LNC=2)

Performance

Config	Throughput	Notes
Prefill (seq=2048)	413 tok/s	Cosine sim 0.99968 vs CPU
Decode BS=32, 1 core	15.6 tok/s	100% token match
Decode BS=32, DP=4	63.0 tok/s	Near-perfect 4.04x scaling

Key Technical Contributions

• NeuronFFT: Pure-PyTorch Stockham FFT replacing torch.fft.rfft/irfft (unsupported on Neuron XLA)
• HBM state management: KV-cache + SSM/FIR state persisted via input_output_aliases across decode steps
• DP=4 process-per-core: Each NeuronCore runs independent inference for 4x throughput

Testing

• 7 files: README.md, src/init.py, src/modeling_evo2.py, test/{init,integration/{init,test_model},unit/init}.py
• Integration tests: NeuronFFT accuracy, prefill compilation + cosine sim, decode block compilation
• All tests validated on trn2.3xlarge with SDK 2.28