Fix QCD Shape unc, calulated at hist merge step, not event wise Unc, so added data-driven flag so that it do not crash before hist-Merge

[Sayan-Dhani]

Title: Exclude data-driven shape uncertainties from per-event tree processing (fixes QCDScale crash)

Description:

Problem
QCDScale is a data-driven shape uncertainty for the HH→bbττ QCD ABCD estimate: its Up/Down templates are built after histogramming, from the OS_AntiIso/SS_Iso/SS_AntiIso sideband regions — there is no per-event weight variation and no corresponding Events__QCDScale__Up/Down tree in the anaTuple.

However, every stage that consumes unc_cfg_dict["shape"] assumed the opposite: that each key in that section names a real tree-level systematic requiring an Events__<unc>__<scale> tree. Adding QCDScale there (to give it a datacard nuisance name) made three independent stages try to load a tree that was never produced, crashing with:

RuntimeError: ERROR: tree Events__QCDScale__Up not found in file ...
KeyError: 'Events__QCDScale__Up'
at HistTupleProducerTask, HistFromNtupleProducerTask, and (latent, not yet hit in production but same code path) AnalysisCacheProducer.

Fix
Introduce a data_driven: true flag on a shape entry (set in the main repo's weights.yaml for QCDScale) and skip flagged entries wherever a stage builds its list of per-event/tree-level uncertainties to process:

Analysis/HistTupleProducer.py (createHistTuple): scale_uncertainties now skips data_driven entries.
Analysis/AnalysisCacheProducer.py (createAnalysisCache): same skip (same bug pattern; would have crashed identically once this stage runs with unc variations enabled).
Analysis/HistProducerFromNTuple.py: uncs_to_compute now skips data_driven entries, so it neither looks up a Events__QCDScale__* tree nor produces {var}_QCDScale_{scale}-named histograms.

The Analysis/HistMergerFromHists.py's fill_hists, which needed the opposite treatment (see HH_bbtautau PR 74 ), and tasks.py's HistMergerTask.uncNames construction, which correctly still includes QCDScale unfiltered — the merge stage is exactly where this data-driven uncertainty is supposed to be built and consumed.