llm: cap decode-program layers so deep small-dim models don't hit the ANE op ceiling

[sbryngelson]

Problem

af.load_llm("HuggingFaceTB/SmolLM2-360M-Instruct") decoded with RuntimeError: execute failed rc=-1. Prefill worked; the failure was the resident-state decode step program (llm.py:601), not the lm_head.

Root cause: _layer_chunks splits the decode program into chunks by baked weight bytes only (_chunk_bytes ~1.6GB). SmolLM2 is deep but narrow (32 layers, dim 960), so all 32 layers stay under the byte cap and compile into one program -- which exceeds an op-count ANE ceiling and fails execute.

I isolated the ceiling with synthetic random models: it trips at ~31 layers regardless of dim (L30 OK / L32 fails at both dim 128 and dim 960), which is why TinyLlama (22 layers) and Qwen2.5-0.5B (24) work while SmolLM2 (32) did not.

Fix

Add _chunk_max_layers = 24 and cap layers-per-chunk by it as well as by bytes. Deep models now segment across decode programs (the machinery already exists; it just wasn't triggered by byte size). 24 sits safely below the measured ~31 ceiling with margin for heavier mixers.

Verification

• Real SmolLM2-360M-Instruct now decodes coherently: "a powerful tool that can help you analyze and visualize complex data..." (was rc=-1).
• Synthetic 32-layer model generates.
• Segmented decode produces the same greedy tokens as a forced single-program run.

Tests (tests/test_decode_chunking.py):

• off-device: _layer_chunks caps at _chunk_max_layers and covers every layer; shallow models stay single-chunk
• on-device (requires_ane): 32-layer model decodes without rc-error; segmented vs single-program greedy match

test_llm.py + new tests: 19 passed, no regressions.