Minimal configs, enable easy CLI, ground work for test suite

[dxqb]

Summary

This PR makes it easy to run multiple presets using a minimal config. It does that by adding two things to train.py:

• --preset-path: applies a built-in preset file before the main --config-path config (without migration, since built-in presets are always saved in the current format).
• --config-value KEY=VALUE: applies a final override after both, with dotted-path support for nested config objects (e.g. ema.decay=0.99). Repeatable.

Example:
./run-cmd.sh train --preset-path "training_presets/#flux2 LoRA 16GB.json" --config-path "test/minimal.json" --config-value output_model_destination=models/my_lora.safetensors

The minimal.json (added in #1545) is, as the name suggests, minimal:

{
    "epochs": 2,
    "batch_size": 2,
    "tensorboard": false,
    "concepts": [
        {
            "path": "resources/icons",
            "text": {
                "prompt_source": "concept",
                "prompt_path": "resources/test/logo.txt"
            }
        }
    ],
    "samples": [
        {
            "prompt": "a mouse eating the OneTrainer logo",
            "width": 512,
            "height": 512,
            "cfg_scale": 3.0,
            "diffusion_steps": 20
        }
    ]
}

The idea is to layer config values on top of each other:

1. defaults as defined in TrainConfig
2. values taken from the preset file
3. values taken from the config file
4. any --config-value overrides

By doing that, your config can be very simple and isn't model-specific anymore.

This isn't a test suite, but it enables one by making it easier to run CLI trainings, @O-J1 @hameerabbasi — see #1545 for the smoke-test scripts built on top of this.

AI assistance

• AI-assisted — I have read every line in this diff and can defend each change