feat(swr): implement ModflowSwr1 write and load

[eacuellarq]

Refs #2389

flopy.modflow.ModflowSwr1 was a stub: write_file() raised
NotImplementedError, load() did not parse the file, and swr was not
registered in Modflow.load. This PR replaces it with a full implementation
of the SWR1 Process (MODFLOW-NWT 1.3.0 / SWR1 1.05).

What's included

• write_file and load for every input item (1–15): options, solver
  controls, reaches and connectivity, tabfiles (ASCII and binary, with the
  time series loaded into tabfiles[i]["data"]), observations, boundary
  types, rain/evap/lateral flow (lists or grids), geometry types 1–5
  (including cross sections and tables), structures of every ISTRTYPE
  (including operable structures and SFR coupling), stage, auxiliary
  variables and direct runoff.
• Where the input instructions and gwf2swr7.f disagree (item 13a field
  order, item 13b, item 4f IOBSLOC2, undocumented options), the code is
  followed; the differences are listed in the module docstring.
• validate() (run by write_file) reports input MODFLOW would stop on or
  silently misread: reach/grid indices, non-symmetric connectivity,
  structures connected to unconnected reaches, TABDATAn/REACHn
  references, missing geometry fields, lines longer than the 200-character
  Fortran buffer.
• SwrExternal keeps lists and grids in OPEN/CLOSE files; loading keeps
  OPEN/CLOSE references.
• "swr" is registered in Modflow.mfnam_packages.
• The constructor stays compatible with the stub (ModflowSwr1(m)).

Index convention

Reach, layer, row and column numbers default to one-based, as in the
SWR1 manual, because reach numbers also appear in connectivity, structures,
observations and REACHn references. zero_based=True makes
KRCH/IRCH/JRCH zero-based like other FloPy packages. Happy to change the
default if preferred.

Tests (autotest/test_swr.py)

• Load of the USGS SWR test problem 5 (added as
  examples/data/swr_sample05, ~350 KB of inputs) and text-identical
  load → write → load → write.
• @pytest.mark.slow: load → write → run of the USGS problem gives
  byte-identical .hds, QAQ, group flow, stage and RIV outputs.
• A channel model built from scratch runs in mfnwt, the flood wave passes,
  and the binary outputs are read with SwrStage/SwrExchange/SwrBudget.
• Every structure type and geometry option written, read by mfnwt, and
  round-tripped.
• Validation errors, option conflicts, zero-based indices, OPEN/CLOSE files,
  binary tabfiles.

Ran locally: ruff check / ruff format --check clean, codespell clean,
pytest autotest/test_swr.py autotest/test_swr_binaryread.py 17 passed.