Add plies-until-progress moves-left head

[mooskagh]

Exposes the per-frame plies_until_progress signal (computed by FillPliesUntilProgress) to training as a second moves-left-style scalar head, and tidies up moves-left target selection.

Data

• Extend the values tensor [6,3] → [6,4] ([q, d, m, p] per value type). Component 3 (p) carries plies_until_progress on the RESULT row only (NaN elsewhere). The 0xff "unknown" sentinel (truncated games / positions after the last progress move) maps to NaN.
• Rename ply_until_progress → plies_until_progress (and FillPlyUntilProgress → FillPliesUntilProgress) for consistency.

Loss config

• MovesLeftLossConfig gains a Component enum (Q=0, D=1, MOVES_LEFT=2, PLIES_UNTIL_PROGRESS=3) selecting the target column, plus required scale (a multiplier, e.g. 0.05) and huber_delta (threshold in plies). Q/D are rejected for moves-left losses; no silent defaults.
• The progress head needs no new model code — MovesLeftHead is reused via config.

NaN masking

• Data losses (MovesLeft/Value/ValueError/ValueCategorical) now zero their contribution per-sample when the target is NaN. Targets are sanitized before the loss so gradients stay finite, then masked. Also fixes latent NaN poisoning from ORIG targets.

Tests / docs

• New model/test_loss_function.py (component selection, Q/D rejection, required scaling, NaN masking with finite grads).
• C++ tensor_generator_test / chunk_rescorer_test updated for [6,4] + the NaN sentinel.
• docs/example.textproto demonstrates a progress head + loss.

All checks green via just pre-commit (clang-format, ruff, mypy, 16/16 C++ tests, 29/29 pytest).

Note: not yet verified end-to-end in the TUI.