Fix the missing PBS progress report, the progress line and stalled-worker detection

[zaihuaji]

Summary

• Missing progress report. check_batch_deadline() was only reached from process_one_backup_file() and transfer_quasar_tarfiles(), i.e. only where a tar was actually dispatched. Every continue above the dispatch skipped it, so a long stretch of skipped files walked past the walltime cutoff without ever parking a report. It is now called from the top of all four per-item loops.
• Misleading progress line. The report led with the tar queue depths, but -A 3 tars each input file right after creating it, so the N count stays near zero exactly while building is the busy phase. New batch_done_count() reports dcount/fcount instead.
• Stalled-worker test. worker_progress_stalled() used an absolute done fraction (MINWDONE = 0.01), so a job at 2% read as progressing and never got a second worker even though it projected to ~28 days. Now rate based: MINWPROJ = 0.9, stalled when done*WALLTIME/elapsed < MINWPROJ.

Test plan

• python3 -m py_compile clean
• 500 all-skipped files and 500 sub-threshold files each hit the walltime guard 500 times (was 0)
• one report per run, none at 0 files, child never reports, guard parks at most once
• batch_done_count() reproduces the live dscheck gc numbers (130927<:>2604 -> 2604 of 130927 file(s) done(1%))
• stall test: production case flips to stalled; grace period, dead process and unlocked-job guards all preserved
• watch the next -A 3 batch run send a progress report with the new line

Generated with Claude Code