Summary

Models trained with sliding-window attention — Mistral 7B, Gemma 3, Gemma 4,
Llama 4 Scout, etc. — only need each layer to attend to the last W tokens.
export_static_llm_coreml.py was always sizing the per-layer KV cache to
max_context_len - input_len, so longer contexts were proportionally more
expensive in both KV cache memory and per-token attention compute even
though the model was trained to ignore everything outside the window.

Add a --sliding_window flag that caps the cache at the trained window.
The downstream pieces — StaticAttentionMask invariants under cache
eviction (validated by test_sliding_window_cache_and_mask) and
StaticAttentionIOManager's per-layer cache_lens plumbing — already
support this; the export script just needed to expose it.

The cache_len computation is factored into _resolve_cache_len so it is
unit-testable. Per-layer mixed sliding/full attention (Gemma 3 / Gemma 4
alternate sliding and full layers) is intentionally left for a follow-up;
this PR uses one window for every layer. Documented in the ANE
Optimizations section of readme.md.

Memory savings example

For a 32-layer / n_kv_heads=8 / head_dim=128 model exported with
max_context_len=8192 in fp16:

Test plan

Added unit tests in examples/apple/coreml/llama/test.py:

• test_resolve_cache_len_no_sliding_window — default path is unchanged.
• test_resolve_cache_len_with_sliding_window — cache shrinks to the window.
• test_resolve_cache_len_sliding_window_larger_than_context_is_a_no_op — a
  user-provided window larger than the remaining context degenerates to the
  no-window case (so users can pass the model's training window verbatim).
• test_resolve_cache_len_rejects_non_positive_window — input validation.
• test_create_example_inputs_with_sliding_window_shrinks_kv_cache — full
  path: every cache tensor in the example inputs has its sequence dimension
  equal to the sliding window, and the attention mask covers
  input_len + sliding_window.

$ python -m pytest examples/apple/coreml/llama/test.py -v
test.py::test_split_model PASSED
test.py::test_resolve_cache_len_no_sliding_window PASSED
test.py::test_resolve_cache_len_with_sliding_window PASSED
test.py::test_resolve_cache_len_sliding_window_larger_than_context_is_a_no_op PASSED
test.py::test_resolve_cache_len_rejects_non_positive_window PASSED
test.py::test_create_example_inputs_with_sliding_window_shrinks_kv_cache PASSED
============================== 6 passed in 5.27s ===============================

I also confirmed the existing examples/models/llama/tests/test_static_attention.py::test_sliding_window_cache_and_mask already covers the cache + mask invariants under both shift_pointer and smart_mask eviction styles when cache_len < total_tokens, so this PR does not need to re-test that.

Authored with Claude.