Add GSPO loss

[jlamypoirier]

Summary

• Adds Group Sequence Policy Optimization (per-segment geometric-mean IS-ratio clipping) as a sibling to GRPO. Sequence-level surrogate, computed as a per-token sum so the softmax-chain backward and SDP partitioning fall out identically to GRPO.
• Extracts shared abstract bases (LanguageModelPolicyGradientLossConfig / LanguageModelPolicyGradientLoss) for the common scaffolding (epsilon_low/high, new_logprobs metric, preprocessing flags) and refactors GRPO to inherit from them.
• Lifts document_index_q/document_index_k from MixerKwargs up to BlockKwargs so the LM head can read them without depending on attention-namespaced keys.
• Renames fast_llm/layers/language_model/loss/grpo.py to policy_gradient.py (now contains both losses + shared base).

Design notes

The GSPO kernel structurally parallels fused_grpo_loss_forward_backward:

1. Same softmax → predicted_logits → new_log_probs setup.
2. New mid-kernel block: scatter-add per-token log_ratio, advantages, and counts into per-segment buffers; optional SDP all-reduce of the three buffers; compute R_s = exp(lrn_sum / token_count) and A_s = adv_sum / token_count (detached).
3. Broadcast R_s, A_s back to per-token; per-token loss weight is mask / token_count_s so each segment contributes once to the sum.
4. Downstream loss and the softmax-chain backward are line-for-line identical to GRPO with probability_ratio = R_{s(t)} and per-token advantages replaced by A_{s(t)}.

The per-token decomposition gets SDP correctness "for free": each rank only sums contributions from its own tokens, so SUM-reducing at the LossDef level reproduces the canonical single-rank result with no /sdp_size correction.

No Triton variant yet — comes in a follow-up.

Test plan

• pytest tests/layers/test_lm_losses.py::test_gspo_loss — 20 cases (10 param sets × 2 batch shapes)
• pytest tests/layers/test_lm_losses.py::test_grpo_loss — 20 cases pass after refactor
• pytest tests/layers/test_lm_losses.py::test_grpo_metrics — 40 cases pass after rename
• End-to-end RL training run (not exercised by this PR)