feat: multi-branch template scanning from git repo + manual template selection

- Rewrite template.rs to scan all remote branches via git commands
  (git fetch/branch -r/ls-tree/git show/git archive)
- Add manual template picker dropdown in CreateForm UI
- Remove sentence-transformers/embed.py from Dockerfile (separate container)
- Clean up Gitea API approach, use local git repo instead
- Add chat panel and sidebar layout improvements

src/api/chat.rs

@@ -0,0 +1,53 @@
+use std::sync::Arc;
+use axum::{
+    extract::State,
+    http::StatusCode,
+    response::{IntoResponse, Response},
+    routing::post,
+    Json, Router,
+};
+use serde::Deserialize;
+
+use crate::llm::{ChatMessage, LlmClient};
+use crate::AppState;
+
+#[derive(Deserialize)]
+struct ChatRequest {
+    messages: Vec<SimpleChatMessage>,
+}
+
+#[derive(Deserialize)]
+struct SimpleChatMessage {
+    role: String,
+    content: String,
+}
+
+pub fn router(state: Arc<AppState>) -> Router {
+    Router::new()
+        .route("/chat", post(chat))
+        .with_state(state)
+}
+
+async fn chat(
+    State(state): State<Arc<AppState>>,
+    Json(input): Json<ChatRequest>,
+) -> Result<Json<serde_json::Value>, Response> {
+    let llm = LlmClient::new(&state.config.llm);
+    let messages: Vec<ChatMessage> = input
+        .messages
+        .into_iter()
+        .map(|m| ChatMessage {
+            role: m.role,
+            content: Some(m.content),
+            tool_calls: None,
+            tool_call_id: None,
+        })
+        .collect();
+
+    let reply = llm.chat(messages).await.map_err(|e| {
+        tracing::error!("Chat LLM error: {}", e);
+        (StatusCode::INTERNAL_SERVER_ERROR, e.to_string()).into_response()
+    })?;
+
+    Ok(Json(serde_json::json!({ "reply": reply })))
+}