feat(hslm): gradient checkpointing for memory efficiency

[gHashTag]

Summary

Gradient checkpointing to reduce memory from O(L) to O(L/checkpoint_every) activations.

New file

• src/b2t/gradient_checkpointing.zig — 170 LOC

Memory savings

• Full: store all L layer activations (e.g., 9 layers × 243-dim × 4B = 8.7KB/batch)
• Checkpoint every 2: store ~L/2 activations, recompute others during backward
• Tradeoff: ~30% more compute for ~50% memory reduction

Features

• CheckpointStore: save activations every N layers, FIFO eviction when full
• MemoryBudget: compute checkpoint interval from total RAM budget
• memoryUsedMB(), savedMemoryMB(): memory tracking
• Configurable: checkpoint frequency, max checkpoints

Tests (5)

• Save and count checkpoints
• Skip non-checkpoint layers (every_n=2)
• FIFO eviction when max reached
• Memory tracking accuracy
• Budget-based interval recommendation

Closes #317