fix: let an unchanged conversation hit the prompt cache

[Javinator9889]

Fixes the first defect in #733.

can_use_cache requires at least two new messages:

const size_t prefix_len = messages.size() - 2;
const bool can_use_message =
    message_checksums_.size() <= prefix_len && matched == message_checksums_.size();

But message_checksums_ is updated to the full incoming list after every
request, so a resend of an unchanged conversation has
message_checksums_.size() == messages.size() and always fails the bound.

That is exactly what a client sends after a timeout. The miss re-prefills the
whole conversation — slower than the request that just timed out — so the retry
times out as well and the loop sustains itself. Observed with OpenCode on a
146-message, ~98k-token conversation.

Bounding by the full incoming length lets the resend hit. The second hunk guards
the case that exposes: with nothing left to prefill, _chunked_insert computes
zero chunks and returns a default-constructed (empty) logits buffer straight into
sampler->sample(). Without it this change crashes the connection.

Verified on qwen3.5:9b: usage contract and generation output unchanged, and a
buried-value probe returns the same string 4/4 on both cold and cached paths.

Note this does not make the fully-cached case cheap — it still re-prefills.
Doing better needs engine-side support; see #733 for the measurements.