fix: keep generated datasets readable for simulation and bootstrap

[nptaylor-irx]

Fixes CSV header parsing failures that blocked simulation and bootstrap during the Sep 16 Astra bug bash. Generated datasets now remain readable when loaded and written into a run folder.

What failed and why

• Michael’s bootstrap treated the quoted header "TAI" as a number. Passing a CSV filename to create_model_from_file() bypassed the unquoted rewrite used for data frames; IGNORE=@ does not skip a row starting with a quote.
• Jordan’s simulations also failed on unquoted headers such as ID and EVID: those models had no character IGNORE rule, so unquoting alone could not fix them.

The fix

A shared writer binds the working CSV and its model’s $DATA record together. It writes unquoted data fields and makes the header match the existing character IGNORE rule, adding IGNORE=# when none exists. Run-folder preparation uses the same writer so it cannot undo the repair.

Existing character and conditional filters, $INPUT DROP flags, column order, and NONMEM null behavior are preserved. Source files are untouched. Blanket IGNORE=@ would broaden some existing filters and silently discard valid rows, so the fix preserves each model’s rule. Package version: 0.0.0.9141.

Validation

Regression tests reproduce both failures and cover CSV/data-frame inputs, filters, DROP columns, source preservation, nulls, and reloading prepared models. Focused tests passed on Pharmpy 2.0 (8 skips: 7 runtime-dependent and 1 explicitly skipped); loader tests passed on Pharmpy 1.8 with no skips. Full NONMEM simulation/bootstrap execution was not run.

After merge, rebuild the pharma_ai_r image with this package version and run simulation/bootstrap smoke tests.

Related: Sep 16 bug-bash feedback.