Build the transformation executor

[swaraj-neu]

The model can already say that a column needs to be restructured, for example reshaped
from wide to long format, but nothing in the code acts on it. This builds the piece that
does.

Scope:

• define the set of transforms the model is allowed to request
• execute them against the existing MSstatsConvert processing functions
• unit tests covering each transform
• fail clearly when a requested transform is not supported

This is the biggest gap in the current implementation and the main engineering
contribution of the project.

Note that once transforms actually run, validation has to check cell values and not only
column names. Worth thinking through alongside the correctness metric.

[swaraj-neu]

Scoping note from looking at the actual data: the executor needs two distinct actions, not one.

split_long handles several values packed into a single cell with a delimiter. DIA-NN Fragment.Quant.Raw is this case, with a variable count per row (7 to 13 observed on Puyvelde2022). This is the action already specified in TRANSFORM_CONTEXT and .transform_entry_type().

A column-to-row pivot handles one measurement spread across many columns. SingleCell-Payne-Custom is this case: 91 columns named Abundances Scaled F1 Sample through F91, one per sample. msstatsconverter.R handles it with pivot_longer. This action is not defined in the prompt at all, so the model has no vocabulary to request it.

The two need different parameters. split_long needs a delimiter and an ID prefix. A pivot needs a column pattern or list, plus where the column name itself maps to (in the Payne case the sample identifier embedded in the column name becomes Run). Worth designing them together rather than extending split_long.