Apply max_kv_size to models that build their own cache

[Lazarus-931]

make_prompt_cache deferred entirely to a model's make_cache(), so max_kv_size was silently dropped on 97 of 173 models.

Verification

e2e on cached checkpoints, a 32-token bound over a ~200-token prompt

model	before	after
Qwen3.5-0.8B	identical output, still answers "Tuesday"	answers "Monday" — context genuinely lost
LFM2.5-1.2B	identical output, still answers "Tuesday"	output changes
Qwen2.5-0.5B (no make_cache)	already worked	unchanged

Across 200 model directories, of the 72 instantiable from a generic config: 44 now apply the bound, 3 were already bounded tighter and are left alone, 1 correctly raises UnboundedKVCache, and none silently ignore it.

[Blaizzy]

Thanks Alazar, this indeed addresses a real issue!

However, these changes look like a hack, basically too much policy in one helper: signature introspection, recursive cache rewriting, status bookkeeping, and a new exception.

We could solve this in a much simpler way.

[Lazarus-931]

would a

-    def make_cache(self):
-        return self.language_model.make_cache()
+    def make_cache(self, max_kv_size=None):
+        return self.language_model.make_cache(max_kv_size)

fix across all model definitions make sense?

[Blaizzy]

Why? When make_cache makes no use of max_kv_size?

qwen3/language.py

def make_cache(self):
   from ..cache import KVCache

   return [KVCache() for _ in self.layers]

qwen3_5/language.py

def make_cache(self):
    return [ArraysCache(size=2) if l.is_linear else KVCache() for l in self.layers]