Fix ArraysCache missing is_trimmable/trim for hybrid model prompt cache

[EagerofLight]

Problem

Hybrid models (Qwen3.5, Qwen3.6, Qwen3-Next) mix ArraysCache (DeltaNet recurrent layers)
with KVCache (attention layers). ArraysCache inherits is_trimmable() → False from
_BaseCache and has no trim() method, so can_trim_prompt_cache() returns False for
the entire hybrid cache.

This silently breaks:

• Prefix cache reuse in mlx_lm.server — fetch_nearest_cache() skips the trim path
• LRU prefix dedup — insert_cache() can't evict redundant prefixes
• Speculative decoding — generate.py raises ValueError on non-trimmable caches

Root cause

# cache.py — _BaseCache default
def is_trimmable(self):
    return False  # ArraysCache inherits this

# cache.py:88 — gate check
def can_trim_prompt_cache(cache):
    return all(c.is_trimmable() for c in cache)  # one False → entire cache rejected

Solution

Add is_trimmable() → True and trim(n) to ArraysCache.

Recurrent state is a compressed summary of all past tokens — it can't be partially
rolled back like KV cache. So trim(n) resets the recurrent state to empty; the
recurrent layers recompute on next forward while KVCache layers still benefit from
their own trim. For exact-match reuse (same prompt repeated), trim() is never
called and the full state is preserved.

Verification (Qwen3.5-4B-4bit, M4 16GB, mlx_lm.server)

Request	TTFT	Cached
Cold (571 tok)	2030 ms	0 / 571
Exact match	314 ms	570 / 571
Prefix reuse	397 ms	424 / 431
Extended prompt	461 ms	564 / 577

Before this fix, prefix reuse and extended prompt got 0 cached tokens.

Fixes #1162