serve: KVCache::new and OptimizedKVCache::new commit num_layers-1 full layers at construction (the form #4778 removes from StreamingKVCache) #4783

crates/aprender-serve/src/inference/kv_cache.rs lines 61-62 (KVCache::new) and 255-256 (OptimizedKVCache::new) build
vec![vec![0.0; max_seq_len * hidden_dim]; num_layers] for both K and V.

The inner zero vector is allocated zeroed and untouched, but vec![x; num_layers] makes the other num_layers - 1 layers
with Clone, which copies them, so their memory is committed before the first token is written.

StreamingKVCache had this form. Three of its memory-bound tests ran out of memory in CI, were ignored under FLAKE-0, and
#4778 makes that allocation lazy.

Evidence. The independent review of #4778 (noted in its PR body, not in its diff). Both sites re-read on main at
75e7b36: they match; the file was last changed in a8c6807.

Fix. The #4778 lazy form, or build each layer separately ((0..num_layers).map(|_| vec![0.0; n]).collect()), so a
layer stays untouched until it is written. Add a test in the style of #4778's.

Scope. No test is ignored for these two types today, so this is not on the 0.70.2 release path. Work starts after
#4778 merges.

Owner aprender-wmainred. Refs #4778.