Add grounded information extraction rewards for GRPO

[ALI-AL-MARJANI]

Motivation

GRPO's reward machinery is domain-agnostic — there's no reason to limit it to math. This
PR adds 5 composable reward functions for hallucination-free information extraction via
GRPO.

Core idea: instead of rewarding correct math answers, reward the model for verbatim
citations. Every extracted_quote must be an exact substring of the source context. This
eradicates factual hallucinations at training time.

Output format

{
  "reasoning_path": "<chain-of-thought>",
  "is_context_sufficient": true,
  "final_answer": "<answer>",
  "extracted_quotes": [
    {"chunk_id": "doc_0", "exact_quote": "<verbatim substring from context>"}
  ]
}

New reward functions

Reward: grounded_format
Signal: Additive JSON schema check — valid JSON → keys → types → quote structure
────────────────────────────────────────
Reward: quote_grounding
Signal: exact_quote ∈ context_raw — core anti-hallucination signal
────────────────────────────────────────
Reward: chunk_routing
Signal: Quote must come from the correct gold chunk, not a distractor
────────────────────────────────────────
Reward: answer_faithfulness
Signal: Token overlap between final_answer and extracted quotes
────────────────────────────────────────
Reward: reasoning_quality
Signal: Encourages substantive CoT

All functions register in REWARD_FUNCS_REGISTRY. Activate via reward_funcs: 
[grounded_format, quote_grounding, ...] in any GRPO recipe YAML — zero changes to the
training loop.

Tests

42 tests, all passing. Covers hallucination detection, correct abstention, partial
grounding, batch processing, multi-chunk routing.