feat: unify shared-prefix Megatron execution

[bradhilton]

Summary

Adds support for deeper shared-prefixes and multi-LoRA slots while simplifying the code. Also adds a dev script for validation and some tests.

What changed

• Replaced depth-specialized shared-prefix/GDN paths with one generic tree representation and execution path.
• Added pack_shared_prefixes(...) and shared-prefix tree/parser coverage for depth 0 through deeper branchy trees.
• Simplified CP/GDN planning and execution around tree metadata, parent indices, and depth buckets.
• Added direct block-mask metadata parity tests against PyTorch create_block_mask, including full/partial block metadata rather than only mask_mod behavior.
• Kept LoRA/vLLM codec/export support aligned with the Qwen3/Qwen3.5 A3B layouts.
• Added dev/megatron_review_perf.py for Austin-focused block-mask, FlexAttention, and CP planning measurements.

Validation

Local on this branch:

• uv run --no-sync prek run --all-files: passed.
• uv run pytest tests/unit/test_shared_prefix_attention_builder.py tests/unit/test_shared_prefix_tree.py tests/unit/test_shared_prefix_packing.py tests/unit/test_shared_prefix_grad_parity.py tests/integration/megatron/lora/test_lora_disk_codecs.py: 30 passed, 9 skipped locally; GPU/Megatron/vLLM portions are skipped on local macOS.

4x H200 SkyPilot (codex-trainer-rank-h200) on this branch:

• Shared-prefix/block-mask/tree/grad-parity subset: 48 passed.
• vLLM LoRA disk codec suite with GPU allocation: 17 passed.
• Austin-style review perf, Qwen/Qwen3.5-35B-A3B-style packed shape, CP=4, 198k packed tokens, 2.448M logical tokens:
  • cold CP planning: 60.54ms
  • warm-cache CP planning: 1.865ms
  • full block-mask build, 7 masks: 136.503ms
  • 8192-token small reference variant: block-mask build 3.478ms, CP planning 12.916ms
• In head-to-head performance check against main, speed was statistically indistinguishable, both about 33k tok/s.