perf: presize paged pool per prefill span and eval grown slabs eagerly

[inureyes]

Summary

Fixes the paged-pool slab growth footprint spikes measured in #224.

• write_prefill sizes a layer's pool once for the whole prefill span: the new presize_for_span counts span blocks that still need a physical row, nets out reusable free rows, and either allocates the pool at the right size (first write) or grows it once. Previously assign_row regrew the slab every POOL_GROW_CHUNK_BLOCKS = 32 rows and every step reallocated and copied the entire layer tensor inside one lazy graph.
• grow_pool evaluates the grown K/V copies immediately, so each old slab returns to the MLX buffer cache before the next layer grows. A multi-layer growth episode now transiently holds one layer's old+new pair instead of every layer's pair.

Measurements (scripts/bench_memory_footprint.py, shared_prefix_burst, M1 Ultra)

case	before	after	dense reference
llama-3.2-1b, ~3.7k-token prefix x8	2815 MiB	2225 MiB	2326 MiB (paged now 0.96x)
qwen3-0.6b, ~4.2k-token prefix x8	8752 MiB	5930 MiB	3833 MiB

The acceptance target (paged within 1.15x of dense) is met on the harness model. The residual qwen3 gap comes from old-slab ladders accumulating in the MLX buffer cache across eight successive pool-extending prefills; that regrowth disappears once cross-request prefix sharing stops re-prefilling shared content (#225, #227), so it is not chased further here. Burst wall times moved 12.6 to 12.9 s (llama) and 21.8 to 23.4 s (qwen3 burst phase), within run noise.

Validation

• cache::paged_pool_tests 25/25 including two new tests: single-step presize for a multi-chunk span (capacity lands at the rounded target), and incremental growth past a presized capacity staying byte-identical.
• Real-model gates, all byte-identical: paged_scheduler_parity (4), paged_prefix_share_parity (2), paged_real_model_parity (1).
• Cold clippy clean, fmt clean. Full mlxcel-core lib suite 865 passed; the single failure (cache::turbo::pack3::tests::packed_bytes_rejects_non_multiple_of_8) is pre-existing on main in release mode (debug_assert-based should_panic) and unrelated.

Closes #224

Review follow-up

Two reviewer findings addressed before merge:

• Test efficacy: the original tests also passed on the old incremental-growth code (final capacity converges identically). A grow_events counter now pins the behavior: 0 slab-copy growths for a presized multi-chunk span, exactly 1 for the growth-past-capacity case. Removing presize_for_span fails the suite.
• Decode-path note: write_prefill also serves single-token decode writes (KVCache::write_paged), so the eager grow eval fires there too, once per 32 x block_size decode tokens per layer in lockstep across layers. It is a bounded, rare sync (and it bounds the transient there as well); scenario wall times stayed within run noise.