Expose remaining training hyperparameters as CLI flags

[will-lamerton]

Problem

nanotune train exposes only --iterations and --lr. Four settings that exist in ConfigSchema have no flag at all:

• batchSize
• numLayers
• stepsPerEval
• saveEvery

So experimenting with any of them means editing .nanotune/config.json between runs.

Separately, mlx_lm lora supports a number of options Nanotune never passes (src/lib/mlx.ts):

• --fine-tune-type (lora / dora / full)
• LoRA rank, alpha, dropout
• --max-seq-length
• --grad-checkpoint
• --val-batches
• a training seed

LoRA rank in particular is one of the most useful knobs for fine-tuning quality, and it's currently unreachable.

Proposal

Two separable pieces:

1. Add flags for the four existing config fields. Mechanical, matches the existing --iterations / --lr pattern in src/cli.tsx and src/commands/train.tsx.
2. Extend the schema and the MLX invocation with LoRA rank/alpha/dropout, --fine-tune-type, --max-seq-length, and a training seed. Larger, needs schema changes and defaults that don't break existing projects.

Worth splitting if someone picks this up.

Acceptance criteria

• Every training.* field in ConfigSchema has a corresponding CLI flag
• CLI flags override config, matching current --iterations behaviour
• New MLX options are schema-backed with sensible defaults
• docs/configuration/index.md and docs/commands/train.md updated

Notes

Part 1 is a good first issue: follow --iterations from src/cli.tsx through to MLXTrainingOptions.

[rohanshrma222]

@will-lamerton, I am working on this issue.