Fix multi-turn crash on hybrid models with untrimmable caches (#2152)

_prefix_cache_trim_amount only returned None (forcing a cold prefill)
when a cache had already evicted part of the prefix. Caches that cannot
be trimmed at all -- ArraysCache, used by hybrid/linear-attention layers
-- slipped through: they are fully retained but have no trim(), so the
per-turn reuse loop's c.trim(n_drop) raised AttributeError on the second
turn of any conversation.

Also require is_trimmable() for every cache, so any non-trimmable cache
forces a cold prefill instead of crashing. Dense and rotating caches are
unaffected (all report trimmable); hybrid models still reuse the cache
when nothing needs dropping (n_drop == 0).