feat(megatron): inherit sweep MBS/GBS/precision from train_params and pass Primus TP/PP

[sukesh-amd]

Summary

• Sweep cells take MBS / GBS / PRECISION from the combination key (and matching body overlays). When sweep is omitted, the implicit default cell requires train_params.micro_batch_size, global_batch_size, and precision instead of falling back to 2 / 128 / BF16.
• Declared combinations keep optional body overlays (training_iterations, tensor_parallelism, pipeline_parallelism); Primus now passes --tensor_model_parallel_size and --pipeline_model_parallel_size.
• Empty sweep.runs with non-empty combinations fails at load. Packaged Megatron JSON and docs match this behavior.

Test plan

• Load a packaged config with a sweep: combo key values win; body overlays apply
• Omit sweep with complete train_params MBS/GBS/precision: one default cell
• Omit those train_params fields with no sweep: load error
• Empty sweep.runs with declared combinations: load error
• Primus run includes TP/PP CLI flags from train_params / combo overlay

[solaiys]

LGTM.

Reviewed sweep-key parsing, implicit default requiring train_params MBS/GBS/precision (no 2/128/BF16 fallback), empty sweep.runs rejected at load, combo overlays, and Primus --tensor_model_parallel_size / --pipeline_model_parallel_size. Packaged JSON and docs match. No blocking issues.