fix(model): make model download pause race-safe

[Jinghao-coding]

中文说明

背景

关联 Issue：Closes #472

模型下载任务在暂停请求与 Reconciler 检查之间存在竞态：Kubernetes Job 可能已被删除，但数据库状态仍未及时更新，导致暂停中的任务被错误标记为失败。

修改内容

• 使用条件状态转换，确保只有 Downloading 状态可以转为 Paused。
• 在暂停事务中释放下载配额。
• 暂停时先持久化 Paused 状态，再删除 Kubernetes Job。
• Reconciler 遇到 Paused 状态时只清理 Job，不再根据 Job 状态覆盖数据库状态。
• Job 不存在时重新读取并加锁，仅将 Pending 或 Downloading 任务标记为失败。
• 增加暂停竞态、暂停任务保留状态和活动任务失败路径的测试。

验证

• make lint
• go vet ./pkg/reconciler ./internal/handler
• 后端 Handler/Reconciler 测试包编译检查
• Controller 和 migrate 二进制构建检查
• git diff --check

本地测试二进制未直接执行，因为测试初始化会连接当前配置的 PostgreSQL 环境；完整测试将由 CI 执行。

English Description

Background

Related issue: Closes #472

There is a race between a model download pause request and the Reconciler: the Kubernetes Job may already be deleted while the database state has not been updated, causing a paused download to be incorrectly marked as failed.

Changes

• Added a conditional state transition so only Downloading downloads can become Paused.
• Released the download quota within the pause transaction.
• Persisted the Paused state before deleting the Kubernetes Job.
• Updated the Reconciler to clean up the Job without overwriting a persisted Paused state.
• Re-read and lock the record when a Job is missing, and only fail Pending or Downloading downloads.
• Added tests covering the pause race, paused-state preservation, and active-download failure paths.

Validation

• make lint
• go vet ./pkg/reconciler ./internal/handler
• Handler and Reconciler test-package compilation
• Controller and migration binary builds
• git diff --check

The local test binaries were not executed because package initialization connects to the PostgreSQL environment configured for this checkout; the complete test suite will run in CI.