Import logic from TaxData repo that creates and tests generation of growfactor.csv file

[martinholmer]

What this branch does

Brings the growth-factors generation logic into TMD from the TaxData repo
(puf_stage1/), so that TMD produces its own growfactors.csv instead of
vendoring a file generated elsewhere. Two commits: one preparing the
make data wiring, one importing the pipeline.

make data changes

• New tmd/create_raw_growth_factors.py reads
  tmd/storage/input/taxdata<VINTAGE>_growfactors.csv and writes
  tmd/storage/output/growfactors.csv.
• tmd/create_taxcalc_growth_factors.py now reads that output file rather
  than reaching into storage/input itself; the TAXYEAR -> input file
  map moved into the new script.
• Makefile: the tmd_growfactors.csv rule runs the new script first;
  logs are split into make_data_raw_growfactors.log and
  make_data_tmd_growfactors.log; clean removes growfactors.csv; the
  data target diffs both new/old growfactors files.

Net effect on make data output: none. The committed
taxdata26_growfactors.csv changes only in line endings (CRLF -> LF);
values are byte-identical otherwise.

New pipeline: tmd/growfactors/

data/ (Census population, CBO baseline, IRS return projection,
       SOI estimates, benefit programs)
  | build_stage1_factors.py     -> data/Stage_I_factors.csv   [level indexes, SYR=2011]
  | build_growfactors.py        -> storage/input/taxdata26_growfactors.csv
  | create_raw_growth_factors.py   -> storage/output/growfactors.csv       [make data]
  | create_taxcalc_growth_factors.py -> storage/output/tmd_growfactors.csv [make data]

• Stage 1 extrapolates SOI aggregates past SOI_YR using Census
  population, CBO baseline, and IRS return-projection growth rates, then
  normalizes to a level index with 2011 = 1.
• Stage 2 converts to per-capita, turns level indexes into
  one-plus-annual-proportional-change factors, appends benefit growth
  factors, and drops variables Tax-Calculator does not use. Takes
  --vintage so a data refresh writes a new vintage file rather than
  rewriting a frozen one.
• tmd/growfactors/Makefile wires the stages: stage1, factors,
  test, all. Most users never run it.
• update/update_cbo.py, update/update_soi.py,
  update/cbo_chained_cpiu.py refresh the source data from live CBO/BLS/
  SSA/IRS sites. Never invoked by make data, make all, or any test;
  gated behind the new pip install -e ".[growfactors-update]" extra
  (requests_html, jinja2) and run by hand via make update-cbo /
  make update-soi.
• Docs: README.md (pipeline overview), docs/annual_update.md (refresh
  procedure), plus the CBO/SOI updating instructions carried over from
  TaxData. .gitignore now un-ignores tmd/growfactors/{data,docs}/*.csv.

Tests

• tests/test_growfactors.py — ported value/range checks on the raw
  growth factors file used for TAXYEAR; runs in the normal make test.
• tests/growfactors_pipeline/test_growfactors_pipeline.py —
  reproducibility tests that re-run each stage into a tmpdir and compare
  against the committed artifact, catching source data edited without
  regenerating downstream files.

Intentional differences from the TaxData original

• Stage_II_targets.csv output dropped — it served only TaxData's stage-2
  weighting; TMD reweights against tmd/storage/input/soi.csv. The
  in-memory aggregates DataFrame (TaxData's Stage_II_targets) is kept
  because stage 1 factors are computed from it.
• Benefit growth factors are padded with 1.0 past the last year in
  benefit_growth_rates.csv; TaxData silently writes empty values for
  2032-2036.
• Two chained assignments that are no-ops under pandas copy-on-write were
  rewritten (see the NOTE comments).
• update_soi.py no longer updates the CPS copy of SOI_estimates.csv,
  which has no TMD counterpart.