Fix MatFormer tier metadata when saving checkpoints

[plugyawn]

Problem

Checkpoint saving currently writes matformer_tier as the effective training tier (effective_tier), not the actual slicing level of the saved weights. When training a full-width checkpoint with --matformer-tier > 0 (runtime slicing), the saved config will set matformer_tier to that tier even though weights are full-width. This makes future auto-detection treat the checkpoint as sliced and can disable runtime slicing or cause incorrect tier handling.

Ref:

• shared/client/src/state/cooldown.rs (writes matformer_tier from effective_tier)

Expected

matformer_tier should represent the actual slicing of the checkpoint weights. If weights are full-width, this should be 0. If weights are sliced, it should reflect the slice ratio.

Possible Approach

• When writing config.json, compute actual_tier from matformer_base_intermediate_size and intermediate_size (if ratio is power-of-two), and set matformer_tier to that value.
• Always store matformer_base_intermediate_size (if not present) to keep detection unambiguous.
• Fall back to 0 when ratios don’t match.

Acceptance Criteria

• Full-width checkpoints always save matformer_tier: 0.
• Sliced checkpoints save the correct tier and base size.
• Auto-detection works for checkpoints saved during training.