fix: correctly size Llama.scores when draft_model implies logits_all

[yadukpb]

What changed

Llama.__init__ already sets self._logits_all = True whenever draft_model is given, regardless of the raw logits_all constructor argument — but self.scores was allocated from that raw argument directly instead of self._logits_all. One-line fix: size self.scores from self._logits_all, consistent with how eval() already uses self._logits_all to decide whether to write into it at all.

Why it matters

Passing draft_model=... without also explicitly passing logits_all=True leaves self.scores sized n_batch instead of n_ctx. eval() writes into it based on self._logits_all, so generation crashes as soon as n_past exceeds n_batch:

ValueError: could not broadcast input array from shape (32000,) into shape (0,)

Found while building a small CPU-only voice agent (potato-agent) that uses LlamaPromptLookupDecoding for speculative decoding — hit this on every turn past ~512 tokens.

#657 reports a superficially similar "could not broadcast ... into shape (0,)" message, but on closer read that one is input_ids overflowing past n_ctx itself (genuine context-window exhaustion) — a different root cause, so I didn't fold this into that thread. #2110 is also draft_model-related but a separate (performance) complaint.

How tested

• git submodule update --init --recursive, make deps (builds the C++ extension from this checkout with the fix applied), make lint — clean.
• Added test_draft_model_scores_buffer_sized_for_n_ctx in tests/test_llama.py: constructs a Llama with draft_model=LlamaPromptLookupDecoding(...) and no explicit logits_all, generates past n_batch tokens. Verified it fails with the exact ValueError above against the pre-fix code and passes against the fix.
• Ran tests/test_llama.py, tests/test_llama_speculative.py, tests/test_llama_chat_format.py, and tests/test_llama_grammar.py (27 tests total) against the from-source build: all pass, no regressions.
• CPU only, Linux x86_64, no special build flags.