fix(sglang): size batch truncation from the longest input in the batch, not the first

[shoemoney]

SGLangModel._greedy_until decides whether to truncate a whole batch from context_size = len(inputs[0]) (src/lighteval/models/sglang/sglang_model.py:271 on main). But GenerativeTaskDataset sorts each split by character length of the query (src/lighteval/data.py:250, -(len(query) + gen_length)), so inputs[0] is the longest prompt in characters, not necessarily in tokens. A batch whose first prompt is short in tokens (long ASCII text) skips truncation entirely, and a later prompt that is shorter in characters but longer in tokens (for example CJK text) is sent to the sglang engine above max_length, corrupting that sample's result.

This is the same defect that issue #1204 confirmed for the vllm backend, whose fix PR #1205 makes its model-side change in src/lighteval/models/vllm/vllm_model.py. Lines 271-292 of the sglang model mirror vllm_model.py:372-397 line for line except for the max_length None guard discussed below, so this PR applies the identical one-line fix to the sglang twin:

context_size = max((len(input_ids) for input_ids in inputs), default=0)

The added unit test builds a two-doc batch where the character-longest prompt tokenizes to 10 tokens and the character-shorter one to 60, with max_length=50 and generation_size=10. On main it fails with 70 not less than or equal to 50 (the 60-token input reaches _generate untruncated); with the fix every input is left-truncated to 40 tokens and it passes. ruff format --check and ruff check are clean.

Deliberately not changed here: unlike the vllm copy, the sglang path has no self.max_length is None guard before the comparison. _create_auto_model defaults _max_length to 8192 when unset, so the None case appears unreachable for sglang, and adding the guard would widen this PR beyond the confirmed defect.