make tiktoken token counting optional via enable_token_counting override

By default, use cheap len/4 estimate for input token counting (metrics
and ratelimit). When enable_token_counting is set to true in overrides,
use tiktoken BPE for exact counts. This eliminates ~80ms of per-request
latency from tiktoken in the WASM filter while keeping metrics and
ratelimit functional.

Made-with: Cursor

crates/common/src/configuration.rs

@@ -131,6 +131,7 @@ pub struct Overrides {
     pub use_agent_orchestrator: Option<bool>,
     pub llm_routing_model: Option<String>,
     pub agent_orchestration_model: Option<String>,
+    pub enable_token_counting: Option<bool>,
 }
 
 #[derive(Debug, Clone, Serialize, Deserialize, Default)]