fix(llm): use --flash-attn on (this llama.cpp build requires explicit value)

The server-cuda image parses -fa as --flash-attn [on|off|auto], so a bare -fa
swallowed the following --cache-type-k as its value and crash-looped.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

stacks/llm/docker-compose.yml

@@ -36,7 +36,8 @@ services:
       - "99"
       - "--ctx-size"
       - "32768"
-      - "-fa"
+      - "--flash-attn"
+      - "on"
       - "--cache-type-k"
       - "q8_0"
       - "--cache-type-v"