feat(channel-monitor): preserve upstream error body

Monitor:
- callProvider now returns both textPath-extracted text and raw body;
  runCheckForModel uses rawBody on non-2xx so history.message stops being
  "upstream HTTP 503: " with empty body (gjson textPath produces "" for
  error responses like {"error":{"message":"No available accounts..."}})
- truncateForErrorBody collapses whitespace then caps at 300 bytes
  (monitorErrorBodySnippetMaxBytes); final truncateMessage still enforces
  the 500-byte DB column cap

Frontend:
- MonitorFormDialog: primary_model input text color and ModelTagInput tags
  now both track form.provider (via new getPlatformTextClass + existing
  getPlatformTagClass with platform prop).

(cherry-picked from 1d3b0418; dropped gateway_handler logging改动，不在本 PR 范围)

backend/internal/service/channel_monitor_const.go

@@ -36,6 +36,10 @@ const (
 	monitorMessageMaxBytes = 500
 	// monitorResponseMaxBytes 单次模型响应最大读取字节，防止 OOM。
 	monitorResponseMaxBytes = 64 * 1024
+	// monitorErrorBodySnippetMaxBytes 非 2xx 响应时保留上游 body 片段的最大字节数。
+	// 留 300 字节足够覆盖典型结构化错误（如 `{"error":{"message":"..."}}`），
+	// 又给 "upstream HTTP <status>: " 前缀留出余量，避免最终被 monitorMessageMaxBytes (500) 截得太狠。
+	monitorErrorBodySnippetMaxBytes = 300
 	// monitorChallengeMin / monitorChallengeMax challenge 操作数范围。
 	monitorChallengeMin = 1
 	monitorChallengeMax = 50