Force a clean prefill after chained KV prefix-cache extensions

[mlpy0]

Long tool loops extend the same prefix-cache entry on every iteration. Those chained partial prefills accumulate numerical drift in the cached KV: on the same prompt, P(end-of-turn) inside an unterminated tool call goes from logprob -11.5 after a clean one-shot prefill to -0.375 once the entry has been extended twice — enough to flip borderline end-of-turn decisions and end the generation in the middle of a tool call.

KVPrefixCache now tracks how many successive deep extensions built each entry's tail and stops offering it for deep reuse at MAX_CHAIN_DEPTH (default 2, EXO_KV_MAX_CHAIN_DEPTH). The next request re-prefills from scratch and the following save resets the entry. Shallow prefix borrows — the shared system/tools block — are unaffected, since that region is written once by the entry's first prefill and never rewritten.

Two smaller fixes from the same incident:

• The master draws a sampling seed when the request doesn't set one, so a retry of a failed generation doesn't replay the exact same failure. Bench requests stay unseeded and deterministic. This is the TODO in mlx_generate.
• The Ollama adapter no longer echoes error_message as assistant content. Ollama's response shape has no error field, so that string is rendered as the model's answer, and it can carry raw generation text. The detail stays in the server logs.

Tests cover the chain-depth accounting (extend / shallow rebuild / shrink / re-save), the reuse and eviction paths at the ratio boundary, and seed defaulting.