server: fix --cache-ram not preventing RAM OOM

[zzhenyao]

Overview

Fix RAM OOM crash in server_prompt_cache when saving prompt cache with --cache-ram limit.

A previous fix added pre-allocation checks, but checkpoint data was not included in the cache size calculation, so the real cache size could still exceed --cache-ram.

Also, update() did not account for the pending allocation, and the states.size() > 1 guard skipped eviction when only one entry existed.

Fixed:

• include checkpoint data in the cache size calculation
• pass pending allocation size and token count to update() so it evicts until the new entry fits
• remove states.size() > 1 guard
• remove catch(bad_alloc) recovery

Additional information

alloc() checked the limit using only KV cache and draft state sizes, and the try block only allocated those. But the cache entry also included checkpoint data, so actual RAM usage could still exceed the configured limit and trigger OOM.

Before fix:

saving prompt with length 146764, total state size = 5332.813 MiB (draft: 307.363 MiB)
oom-kill: ... task=llama-server,pid=92787
Out of memory: Killed process 92787 (llama-server) ...

After fix:

saving prompt with length 72820, total state size = 2721.372 MiB (draft: 152.505 MiB)
single state (2721.372 MiB) exceeds cache limit (360.000 MiB), skipping cache

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, llama.cpp + claude code, used for checking code style

[zzhenyao]

@aldehir @ggerganov could you please take a look when convenient?
Possibly related: #22925, #22629, #21690

Follow-up: if --cache-ram is supposed to include checkpoint too, then this fix should stop prompt cache from exceeding the limit and causing RAM OOM. But that also means the default value may now be too small. Should it be adjusted?