Handle the DIA-NN packed column case

[swaraj-neu]

DIA-NN can put several values in one cell separated by a semicolon, or spread a single
measurement across several columns. The standard processing functions do not handle
either case.

Exploratory. If it is not solved, write it up as an open problem in the final report
rather than forcing a fix.

[swaraj-neu]

Confirmed on the Puyvelde2022 DIA-NN report
(/projects/VitekLab/Data/MS/Benchmarking/DIA-Puyvelde2022-HYEtims735_DIA/HYEtims735_DIA/DIANN/report.tsv).

Fragment.Quant.Raw and Fragment.Quant.Corrected both pack multiple values per cell, semicolon separated. Count varies by row, across 2000 rows sampled: 7 (338 rows), 8 (18), 9 (36), 12 (36), 13 (1572). So a split has to be dynamic, not fixed width. Fragment.Correlations carries the same count and can be used as a consistency check.

What this blocks. Mapping accuracy (#7) is unaffected, since that only compares column names against ground truth. The downstream comparison (#8) is blocked for DIA-NN only: Intensity would point at a packed column, as.numeric() returns NA, and the pipeline yields a table with no intensities and no error. If allow_transforms is on, the field is nulled and NA-filled instead, same outcome. The hand-written DIANNtoMSstatsFormat unpacks internally, which is why the reference path works, so the two paths aren't comparable until the LLM path can do the same.

Scope of the fix. Not the full transform executor. The prompt already specifies action: "split_long", delimiter, id_field, and id_prefix; .transform_entry_type() already exists; the model already emits it. What's missing is reading llm_mapping$transforms in LLMtoMSstatsFormat and expanding one row into N with generated FragmentIon labels.

Other formats. Spectronaut is already long (F.PeakArea is single-valued, with one row per fragment), and the PD fixture (tinytest/raw_data/PD/pd_input.csv) is also already long (one row per PSM, with Sequence, Protein.Group.Accessions, Charge, Intensity, Spectrum.File). So DIA-NN is the only benchmark format needing a transform. Payne (SingleCell-Payne-Custom) is a separate 91-column wide-to-long problem, relevant to #14/#17 but not the pre-paper benchmark.

Also noting Ms1.Area (column 43) is a clean single-value intensity in the same file. A model could map it, pass every structural check, and produce scientifically wrong DIA quantification. Worth capturing in the correctness definition in #6.