server: make exact prompt-cache hits generation-safe

[deyi2026]

Problem

When a client resends the exact same full prompt, LRUPromptCache.fetch_nearest_cache
returns a legitimate exact hit with rest=[]. The empty segment then reaches
BatchGenerator.insert_segments and crashes the request (it needs at least one
prompt token to produce the first logits).

Fix

_generation_safe_cache_hit guarantees generation always has at least one
prompt token after a cache hit:

• trimmable KV: trim one cached token and replay the final prompt token
• hybrid recurrent caches (not trimmable as a whole): fall back to the nearest
  cache for prompt[:-1] and replay the final token
• if no shorter cache exists: recompute the whole prompt (fail-safe)

Non-exact-hit paths are unchanged.

Testing

tests/test_server_cache_exact.py covers both cache kinds and the recompute
fallback (11 cases).