[Code scan] Avoid mutating HDFS backward file lists during download

[njzjz]

Found by a Codex global repository scan of deepmodeling/dpdispatcher at commit 98a9e08.

Problem
HDFS download(back_error=True) reuses the submission's backward file lists directly and appends absolute glob() results to them.

Relevant code

dpdispatcher/dpdispatcher/contexts/hdfs_context.py

Lines 149 to 158 in 98a9e08

	for task in submission.belonging_tasks:
	local_job = os.path.join(self.local_root, task.task_work_path)
	remote_job = os.path.join(gz_dir, task.task_work_path)
	flist = task.backward_files
	if back_error:
	errors = glob(os.path.join(remote_job, "error*"))
	flist.extend(errors)
	for jj in flist:
	rfile = os.path.join(remote_job, jj)
	lfile = os.path.join(local_job, jj)

dpdispatcher/dpdispatcher/contexts/hdfs_context.py

Lines 187 to 195 in 98a9e08

	local_job = self.local_root
	remote_job = gz_dir
	flist = submission.backward_common_files
	if back_error:
	errors = glob(os.path.join(remote_job, "error*"))
	flist.extend(errors)
	for jj in flist:
	rfile = os.path.join(remote_job, jj)
	lfile = os.path.join(local_job, jj)

Impact
The submission object is mutated during download. Later downloads can try to fetch stale temporary paths. Because os.path.join(local_job, absolute_path) ignores local_job, error files may be moved to unintended absolute locations or not copied under the task/common output directory.

Suggested fix
Build local copied lists, for example flist = list(task.backward_files), and append relative paths or basenames for generated error files. Do the same for submission.backward_common_files.