Fix cache block size for flash decoding (#2351)

* Fix cache block size for flash decoding

This seems to have been accidentally dropped during the TRT-LLM
PR rebase.

* Also run CI on changes to `backends`

.github/workflows/ci_build.yaml

@@ -10,6 +10,7 @@ on:
     paths:
       - ".github/workflows/build.yaml"
       - "integration-tests/**"
+      - "backends/**"
       - "server/**"
       - "proto/**"
       - "router/**"

backends/v3/src/backend.rs

@@ -35,9 +35,16 @@ impl BackendV3 {
         window_size: Option<u32>,
         speculate: u32,
     ) -> Self {
+        let flashdecoding = if let Ok(flashdecoding) = std::env::var("FLASH_DECODING") {
+            matches!(flashdecoding.to_lowercase().as_str(), "1" | "true")
+        } else {
+            false
+        };
+        let block_size = if flashdecoding { 256 } else { 16 };
+
         let queue = Queue::new(
             requires_padding,
-            16,
+            block_size,
             window_size,
             speculate,
             max_batch_total_tokens,