fix(training): correct rank-zero log messages, Print total model size once at startup (rank-0) – Fixes #123

[MagellaX]

• After setup_model_and_optimizer returns the model, compute total parameters with the existing get_parameters_in_billions utility.
• Print a single formatted line on args.rank == 0.
• No functional or performance impact—purely a log addition.

Implementation details
Leverages get_parameters_in_billions(model) to keep counting logic consistent.
Converts billions → exact count (int(total_params_B * 1e9)) for clarity.
flush=True ensures the line appears promptly even in buffered environments.
Guarded by args.rank == 0 so it prints once per job, regardless of model/data/pipeline parallelism.

Testing
Unit tests
pytest -q tests → all green.
Smoke-test (examples/pretrain_gpt_tiny.sh)
Confirmed the new line appears once on rank 0 and nowhere else.
Training proceeds unaltered.

Backward compatibility
None of the existing outputs were removed or changed; only one new line is added.