fix: isolate legacy Tinker branch contexts

[bradhilton]

Agent: Tarkin

Part of #1111 (item 9).

Problem: The legacy Tinker service splits each packed row into one datum per branch. It built each branch's mask as (group == g) | (parent == parent(g)). Prefix-tree roots have parent_id == group_id, so that mask pulled sibling branches into every datum's context, and the root's datum covered the whole row. Targets were also shifted across the whole packed row before masking, so a branch's first target could be a sibling's token.

Behavior after this change: Each datum now contains exactly its branch and that branch's ancestors, read from the packer's own prefix-tree segments (parse_prefix_tree_row). Targets are shifted within each datum. Optimizer stepping is unchanged: ART's gradient accumulation is fixed at one packed sequence since #808, so there is still one optimizer step per packed row.

Validation:

• New tests: tests/unit/test_tinker_service_training.py runs the real packer and loss callback against a fake Tinker client. It covers two siblings, a deeper tree, separate roots and padding, checking each datum's inputs, targets and weights, plus the per-row optimizer count. 3 of its 4 tests fail on main.
• Existing tests: the Tinker service test suites pass (58 tests).

Limitations: Tested against a fake client only, with no live Tinker run. Prompt-only root datums are still sent, as before.