feat(torchtitan): enable dynamic model parameter override via CLI

[Xiaoming-AMD]

Summary

This PR adds support for dynamically overriding TorchTitan model parameters from the CLI
(e.g., --model.n_layers=4) during Primus training.

Key Changes

• Added _split_known_unknown() in parser to separate known and unknown overrides.
• Forward unknown overrides (like model.*) to TorchTitan trainer as extra_args.
• Added patch_titan_train_spec() to modify TorchTitan's model configuration dynamically.
  • Supports nested form {"model": {"n_layers": 4}}
  • Enforces strict "model." prefix validation
  • Raises clear errors for missing or invalid fields
• Megatron trainer explicitly rejects unregistered overrides.

Example Usage

EXP=examples/torchtitan/configs/MI300X/llama3.1_8B-BF16-pretrain.yaml bash examples/run_pretrain.sh  --model.n_layers=4

Motivation

Previously, TorchTitan's model configuration was static and not exposed externally.
This patch allows Primus users to quickly experiment with model structure or scale
without editing Titan source configs.

[wenxie-amd]

LGTM