Fix DetailsLogger compatibility with xxhash 4.0

[Honorllq]

Summary

• encode logging values as UTF-8 bytes before passing them to xxhash.xxh64
• use the same byte-safe helper for per-sample, per-task, and cross-task hashes
• declare xxhash as a direct dependency
• add a regression test that simulates bytes-only xxhash behavior with non-ASCII input

Why

xxhash 4.0.0 no longer accepts Python strings directly. DetailsLogger.log() and DetailsLogger.aggregate() passed strings to xxh64, which raised TypeError during evaluation logging.

Closes #1330.

Validation

• targeted compatibility check with xxhash==4.0.0
• python -m py_compile src/lighteval/logging/info_loggers.py tests/unit/logging/test_evaluation_tracker.py
• ruff check src/lighteval/logging/info_loggers.py tests/unit/logging/test_evaluation_tracker.py
• ruff format --check src/lighteval/logging/info_loggers.py tests/unit/logging/test_evaluation_tracker.py

[adhabnr-ux]

Independent confirmation this is worth merging: hit the exact same crash (identical traceback line, DetailsLogger.log() → xxhash.xxh64(doc.query)) from a completely different angle — building an EvalPort adapter that reads real per-document Pipeline.get_details() output from lighteval 0.13.0. Reproduced from scratch on a totally fresh pip install lighteval (today's PyPI resolves xxhash to 4.0.1, same as the original report), confirmed the standalone repro (xxhash.xxh64("hello") alone raises the same TypeError, no lighteval involved), and confirmed pinning xxhash<4.0 unblocks a real Pipeline.evaluate() run end-to-end.

Worth noting for triage: it's not just log() — every one of the four xxhash.xxh64(...) call sites in info_loggers.py (hash.example, hash.input_tokens, hash.cont_tokens, and the three compiled_hash.* aggregation lines) has the same str-not-bytes issue, so a fix needs to cover all four, not just the one in the traceback — sounds like that's already what this PR does per the summary ("per-sample, per-task, and cross-task hashes"), just flagging in case it helps review. Since I had to work around this myself to get real test data (pinning xxhash<4.0 in my own project's test deps as a stopgap), happy to see this land — it currently breaks Pipeline.evaluate(), which is the method every CLI backend (accelerate/vllm/endpoint/etc.) calls under the hood, for anyone installing lighteval fresh today.