fix(cache): treat empty paged length list as a benign no-op

[inureyes]

Summary

During batched speculative decode the Failed to sync paged state for seq-N: CachePool: expected 32 paged layer lengths for seq-N, got 0 warning spammed the server log hundreds of times per burst. The underlying Err was benign: an empty per-layer length list is the model-owned "nothing to mirror yet" case, not a real length divergence. This makes the empty case a legitimate Ok(()) no-op so the spam disappears while a genuine mismatch still errors.

Root cause

The model-owned paged families (gemma3/gemma4/llama4/qwen3_5/qwen3_next) build the per-layer target_lens for CachePool::sync_paged_state_with_lengths from their per-sequence sequence_caches (src/models/gemma3.rs, llama4.rs, qwen3_5.rs). A just-admitted or speculative/draft row has empty sequence_caches before its first model-owned forward, so visible_lens is an empty vec and an EMPTY length list reaches the sync. The length-count guard in sync_paged_state_with_lengths treated 0 != state.layers.len() as a divergence and returned the "expected N ... got 0" error, which scheduler.rs swallowed into a warn! per tick per row. The model-owned paged family uses the paged backend for SHADOW block-table accounting only, so with no populated caches there is genuinely nothing to mirror on that tick.

What changed

• src/lib/mlxcel-core/src/cache.rs: in sync_paged_state_with_lengths, add an early return Ok(()) when target_lens.is_empty(), before the target_lens.len() != state.layers.len() count check, with a comment explaining the model-owned "nothing to mirror yet" no-op and why it is distinct from a nonzero partial count. The nonzero-mismatch error path is unchanged; a real desync still returns Err.
• The fix lives in the shared CachePool method so every model-owned family is covered by one change. The model-owned sync_sequence_storage impls, the scheduler.rs swallowed-Err warn! (which correctly fires only on a genuine Err now), and the paged block-table accounting are untouched.
• Added a CachePool unit test sync_paged_state_with_lengths_empty_is_benign_noop: a 2-layer model-owned shadow state accepts an empty length list as an Ok no-op that leaves the layer lengths unchanged, while a nonzero-but-wrong count still returns the mismatch Err. Pure CachePool unit test, no model load.

Test plan

• cargo test -p mlxcel-core --lib --features cuda sync_paged_state_with_lengths_empty_is_benign_noop (pass)
• cargo clippy -p mlxcel-core --lib --features cuda -- -D warnings (clean)
• cargo test -p mlxcel-core --lib --features cuda cache single-threaded: the new test and all sync/paged tests pass; the only 2 failures (cache::paged_detach::tests::detached_paged_set_accounts_real_pool_bytes, cache::paged_detach::tests::trim_detached_paged_to_enables_partial_prefix_adoption) are pre-existing on clean main (identical assertion failed: handle.offset >= target_tokens in paged_detach.rs, an unrelated module) and are not touched by this change.

Closes #823