Eliminate the CSV intermediary step for GUL, IL/FM, and RI input files.
Previously the pipeline wrote DataFrames to CSV, then converted CSV to
binary at runtime. Now binary files are written directly from DataFrames
during preparation, with CSV output only when intermediary_csv=True
(for debugging).

Preparation changes:
- Refactor csvtobin converters to decouple CSV reading from binary
  writing: extract df_to_ndarray(), amplifications_write_bin(), and
  complex_items_write_bin() as reusable functions.
- write_gul_input_files(): replace per-column pop/prepare dicts with a
  unified files_write_info dict that drives binary+CSV output.
- get_il_input_items(): write FM binary files directly, rename pandas
  dtype dicts to avoid shadowing numpy dtypes, return .bin paths.
- write_files_for_reinsurance(): write RI binary files directly using
  df_to_ndarray().tofile().
- Thread intermediary_csv parameter through GenerateFiles, RunExposure,
  and all preparation functions.

Execution changes:
- Replace csv_to_bin/\_csv_to_bin with move_bin() which relocates
  pre-built .bin files to the run output directory.
- Update IL/RI detection in GenerateLossesDir and
  GenerateLossesDeterministic to check for .bin files (not just .csv).
- Update _check_each_inputs_directory to accept either .csv or .bin.
- Remove unused step_flag from deterministic loss commands.
- Use np.memmap(mode='r') instead of np.fromfile in load_as_ndarray
  and load_as_array for demand-paged binary reads.

Test updates:
- Remove CsvToBin test class and associated test helpers.
- Update test_generate_files expected paths from .csv to .bin.