feat: add performance.yaml example and enable gradient_checkpointing in borealis_1b (issue #6 items 6-7)

[hai-pilgrim]

Summary

Closes parts of #6 (items 6 and 7):

• New examples/performance.yaml: A single annotated config showcasing all high-impact performance knobs:

  • gradient_checkpointing: true — saves VRAM, enables larger batches
  • optim: "adamw_bnb_8bit" — 8-bit Adam via bitsandbytes, cuts optimizer memory ~75%
  • attn_implementation: "flash_attention_2" — 2-4x attention speedup on long sequences
  • per_device_train_batch_size: 4 + gradient_accumulation_steps: 2 — effective batch 8
  • target_modules: all-linear — targets every linear layer for maximum adapter coverage
  • Each knob has an inline comment explaining the tradeoff

• examples/borealis_1b.yaml: Enable gradient_checkpointing: true (was false). For a 1B model on 24GB VRAM, the memory savings allow larger batch sizes that more than recover the ~20-30% wall-clock overhead noted in the issue.

Remaining issue #6 items

Item	Status
1. Unsloth	Done in #14
2. Flash Attention	Done (attn_implementation field + examples already used sdpa)
3. Vectorize logit loop	Done in #20
4. Batch trajectory generation	Done in #17
5. Pre-compute prompt lengths	Already cached via _prompt_length_cache
6. Gradient checkpointing defaults	This PR
7. 8-bit optimizer	This PR (in performance.yaml)
8. Move tokenization to collator	Not yet

Test plan

• All 51 non-benchmark tests pass
• examples/performance.yaml is valid YAML
• borealis_1b.yaml now has gradient_checkpointing: true

🤖 Generated with Claude Code

[marksverdhei]

Review

The performance.yaml example is well-documented and the gradient_checkpointing change in borealis_1b makes sense. Good addition.

Needs rebase: This PR includes the bitsandbytes skip guard that was already merged via #21. Please rebase onto main to drop the duplicate diff — the rest of the PR is clean and ready to merge after that.

[marksverdhei]

Reviewed. The performance.yaml example is well-documented with clear comments explaining each knob. The gradient_checkpointing flip in borealis_1b is appropriate for 1B+ models. The bitsandbytes skip guard overlap with #21 resolved cleanly. LGTM.