implements phase 1 of zero gen verification

[entrpn]

Description

Implements masked_prefill_logits_checker.py, which verifies the architectural and mathematical parity
between a MaxText model implementation and a HuggingFace PyTorch reference model without requiring generation.

Tests

llama3.1-8b:

--- Running Zero-Generation Verification on 2048 tokens ---
INFO:absl:=== Sequence composition ===
INFO:absl:Total tokens (padded): 2048
INFO:absl:Real tokens (unpadded): 1879
INFO:absl:Padding tokens: 169
INFO:absl:Tokenizer padding_side: right
INFO:absl:Batch determinism verified. Max diff: 2.3041e-03, Mean diff: 2.7256e-07, Tolerance: 1.8460e-02 (rel=0.0005 × max_logit=36.92)
INFO:absl:Comparing top-k at last real token position: 1878
INFO:absl:
--- MaxText model top 10 tokens ---
[transformers] Ignoring clean_up_tokenization_spaces=True for BPE tokenizer TokenizersBackend. The clean_up_tokenization post-processing step is designed for WordPiece tokenizers and is destructive for BPE (it strips spaces before punctuation). Set clean_up_tokenization_spaces=False to suppress this warning, or set clean_up_tokenization_spaces_for_bpe_even_though_it_will_corrupt_output=True to force cleanup anyway.
INFO:absl:| Token ID   | Token                | Score      |
|------------|----------------------|------------|
| 128001     | <|end_of_text|>      | 13.4656    |
| 791        | The                  | 12.0378    |
| 74694      | ```                  | 11.5461    |
| 14196      | ``                   | 11.4824    |
| 9468       | �                    | 11.4151    |
| 14924      | Question             | 11.1182    |
| 2          | #                    | 10.5820    |
| 2028       | This                 | 10.4056    |
| 16856      | Testing              | 10.2497    |
| 29184      | James                | 10.2277    |

INFO:absl:
--- HF model top 10 tokens ---
INFO:absl:| Token ID   | Token                | Score      |
|------------|----------------------|------------|
| 128001     | <|end_of_text|>      | 13.4904    |
| 791        | The                  | 12.0613    |
| 74694      | ```                  | 11.5677    |
| 14196      | ``                   | 11.5009    |
| 9468       | �                    | 11.4166    |
| 14924      | Question             | 11.1601    |
| 2          | #                    | 10.5955    |
| 2028       | This                 | 10.4204    |
| 58         | [                    | 10.2445    |
| 29184      | James                | 10.2405    |

INFO:absl:
--- Similarity Metrics of Top Tokens ---
INFO:absl:| Metric                         | Value                |
|--------------------------------|----------------------|
| overlap_count                  | 11                   |
| jaccard_similarity             | 0.8182               |
| rank_agreement_percentage      | 90.00%               |

INFO:absl:Comparing logits over tokens [0:1879] (padding_side=right, real_token_count=1879, padded_total=2048)
INFO:absl:
--- Tier 1B: Logit Parity Metrics ---
INFO:absl:L-infinity Norm (Max Abs Diff): 2.7225e-01 at index (269, 1302)
INFO:absl:Cosine Similarity (Mean):       1.0000
INFO:absl:Cosine Similarity (Min token):  0.9999
INFO:absl:  Threshold check (Mean): 1.0000 > 0.999
INFO:absl:KL Divergence (Max token):      7.8836e-04
INFO:absl:KL Divergence (Mean):           9.0536e-06
INFO:absl:  Threshold check: 7.8836e-04 < 0.5
INFO:absl:✅ Model parity verified successfully
INFO:absl:  Mean cosine similarity: 0.999985
INFO:absl:  Max L∞ diff: 2.7225e-01
INFO:absl:  Mean KL divergence: 9.0536e-06

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.