Resolve the premix file list once instead of in every job

[kandrosov]

Stacked on #16 — until that merges this PR also shows its four commits. Merge #16 first, then this.

What this fixes

All 500 failures of the 5000-job production were one error:

NoSecondaryFiles: RootEmbeddedFileSequence no input files specified for secondary input source
  Constructing module: class=PreMixingModule label='mixData'

--pileup_input dbs:<dataset> makes every job resolve the premix dataset — 37 993 files — with
its own DAS query. At a few thousand concurrent jobs the queries start coming back empty, cmsDriver
writes a config with no secondary input, and cmsRun dies after the job has already produced its
GEN-SIM. The damage was concentrated where outbound access is poorest: T1_IT_CNAF 159,
T2_TR_METU 135, T2_EE_Estonia 60, T1_US_FNAL 34. At 115 100 jobs it gets worse, not better.

How

PremixFileList, a local per-era task, resolves the dataset once and stores the plain file list at
<output>/premix/<era>.txt on fs_default. RunProd requires it, localizes it like the gridpack,
and passes --pileup_input filelist:<path>.

filesFromList (Configuration/Applications/python/ConfigBuilder.py:104) reads one LFN per line —
exactly dasgoclient output — so the job gets precisely what a successful DAS query would have
produced. Nothing about the physics changes; 115 100 DAS queries become five.

Details worth knowing:

• Stored in the production area, never <output>_test: the list depends only on the era, so a test
  run reuses it (same rule as gridpacks).
• A batch node refuses to build one (the submitted_task_family() guard from Do not let a production job build its own gridpack #11) — resolving
  it on a worker is the exact thing being eliminated. Submitting PremixFileList itself still
  works, so lists can be pre-built on the grid.
• With no list configured the command falls back to dbs: unchanged, so no existing setup breaks.

Cost

The list is 5.7 MB, downloaded per job, against the ~29 MB gridpack each job already pulls — under
20 % on top of existing per-job input. A deterministic per-job slice would cut it to near zero but
changes which premix events each job sees, so that is left as a deliberate choice rather than made
silently here.

Testing

Ten unit checks: the dataset is discovered from the conditions for all five eras; without a list the
command still says dbs:; with one it says filelist: and contains no dbs: anywhere; a step
without pileup gets no --pileup_input; the task branches once per era; a test run resolves to the
production path; a worker running RunProd refuses while PremixFileList itself proceeds.

End to end on Run3_2023 (--test 5, local), all green:

PremixFileList[Run3_2023]: 37993 files from /Neutrino_E-10_gun/Run3Summer21PrePremix-...
--pileup_input "filelist:/tmp/kandroso/luigi-tmp-034755548.txt"
Scheduled 9 tasks — 6 ran successfully, no failed tasks
staging/nanoAOD_v12/.../nano_v12_1.root  2 869 283
staging/nanoAOD_v15/.../nano_v15_1.root  3 211 983

The stored list is 37 993 lines / 5 667 589 B — matching DAS's nfiles for the dataset exactly —
and DIGIPremixHLT read the premix events from it, which is the point of the change.

mkdocs build --strict passes; run_tools/apply_format.sh is a no-op on the diff.