feat: add distillation-specific checkpoint management

Add dedicated checkpoint management for knowledge distillation training without
polluting IFullModel with additional interface requirements.

## Design Decision

**Question**: Should we add ICheckpointableModel to IFullModel?

**Answer**: No - create specialized checkpoint manager instead

**Reasoning**:
1. IFullModel already inherits 7 interfaces - adding more violates Interface Segregation
2. Not all models support checkpointing (stateless models, non-serializable state)
3. IModelSerializer already provides similar file-based functionality
4. Distillation-specific needs (curriculum state, teacher/student pairs) don't belong in IFullModel

## New Components

### DistillationCheckpointConfig
Configuration for checkpoint management:
- Save frequency (every N epochs or batches)
- Keep best N checkpoints (automatic pruning)
- Which models to save (teacher, student, or both)
- Curriculum state preservation
- Metric tracking (validation loss, accuracy, etc.)

### DistillationCheckpointManager<T>
Handles checkpoint operations:
- Automatic saving based on schedule
- Best checkpoint selection by metric
- Multi-stage distillation support (student → teacher)
- Curriculum learning state management
- Training resumption after interruptions
- Batch-level checkpointing for long epochs

### Key Features

1. **Automatic Checkpoint Pruning**
   - Keep only best N checkpoints based on validation metric
   - Save disk space while preserving important models

2. **Multi-Stage Distillation**
   - Save student from Stage 1
   - Load as teacher for Stage 2
   - Support progressive compression (large → medium → small)

3. **Curriculum State Preservation**
   - Save curriculum progress with checkpoint
   - Resume from correct curriculum stage
   - Don't restart from "easy samples"

4. **Flexible Checkpointing**
   - Epoch-based or batch-based
   - Stream-based via ICheckpointableModel
   - File-based via IModelSerializer
   - Custom metrics for best checkpoint selection

## Usage Example

```csharp
var config = new DistillationCheckpointConfig
{
    CheckpointDirectory = "./checkpoints",
    SaveEveryEpochs = 5,
    KeepBestN = 3,
    SaveStudent = true
};

var manager = new DistillationCheckpointManager<double>(config);

for (int epoch = 0; epoch < 100; epoch++)
{
    // Train...
    double validationLoss = Evaluate(student);

    manager.SaveCheckpointIfNeeded(
        epoch: epoch,
        student: student,
        metrics: new Dictionary<string, double> { { "validation_loss", validationLoss } }
    );
}

// Load best checkpoint
manager.LoadBestCheckpoint(student);
```

## Documentation

Added CHECKPOINTING_GUIDE.md with:
- Why checkpointing matters for distillation
- Architecture overview
- 5 detailed usage examples
- Best practices
- Common patterns (early stopping, multi-stage, resumption)

## Benefits

- ✅ Separation of concerns (checkpointing separate from model interface)
- ✅ Distillation-specific features (curriculum state, teacher/student pairs)
- ✅ Production-ready (automatic pruning, resumption, multi-stage support)
- ✅ Flexible (works with any ICheckpointableModel)
- ✅ Doesn't pollute IFullModel interface

Resolves checkpoint management requirements for knowledge distillation

Author: claude