Fix prompt cache saving and chat-persistent rollover

[ejones]

Fixes #1670, by reworking the original fix for #1585 from #1609.

The original fix examined embd to determine if the prompt had been evaluated, but embd is limited to the batch size. In addition, that fix left session_tokens in its original state (i.e., the longer, cached prompt), while normal session evaluation truncates it at the first eval. This combination meant that any prompts with a cache hit on just the first batch (512 by default) would begin eval-ing ~from the second batch, and all of that eval would get appended to the end of the full, original cached prompt. This had the downstream effect of diverging the cache from the prompt and overrunning the context size in the cache, as seen in #1670.

For the fix, I opted to move the re-eval logic to main's initialization rather than at the eval stage. Here, it transforms session_tokens such that it will only match (prompt - 1) tokens.

Testing:

• for chat-persistent.sh not rotating cache files correctly #1670, conducted a long chat with 30B, past the context rotation
• for Bug when prompt stored in --prompt-cache is longer than the new one #1585, applied the Z/joke test and got a joke that did not start with "Z"

[github-actions]

clang-tidy made some suggestions

[DannyDaemonic]

This is a clever fix. Feel free to merge after the suggested size() to !empty() fix.

[ejones]

Thanks!