fix: keep sampling args per token

[samsja]

Summary

• Keep rollout sampling args per token by carrying completion_temperatures, completion_top_ks, and completion_top_ps through TrainingSample / MicroBatch.
• Allow packed microbatches to concatenate samples with different sampling args while preserving per-token replay values in the trainer.
• Add exact top-k replay support to the chunked LM head path, including the Gemma softcapped variant; chunked heads still reject top_p < 1.0.

Stacked on #2601.

Verification

• uv run pytest tests/unit/orchestrator/test_batch.py tests/unit/orchestrator/test_trajectories.py tests/unit/orchestrator/test_sft_trajectories.py tests/unit/orchestrator/test_teacher_logprobs.py tests/unit/train/rl/test_packer.py tests/unit/train/rl/test_fused_lm_head.py -q
• uv run ruff check src/prime_rl/transport/types.py src/prime_rl/orchestrator/trajectories.py src/prime_rl/trainer/batch.py src/prime_rl/trainer/rl/data.py src/prime_rl/trainer/rl/loss.py src/prime_rl/trainer/rl/train.py src/prime_rl/trainer/rl/packer.py src/prime_rl/trainer/models/layers/lm_head.py src/prime_rl/trainer/models/layers/lm_head_gemma.py tests/unit/orchestrator/test_batch.py tests/unit/orchestrator/test_trajectories.py tests/unit/orchestrator/test_teacher_logprobs.py tests/unit/train/rl/test_packer.py tests/unit/train/rl/test_fused_lm_head.py
• git diff --check
• Direct chunked top-k sanity probe against full-logits replay