Server: chat/completions live tool continuation via tool_call_id

[starforge-labs]

Fixes #609 - On the Anthropic endpoint, tool rounds continue the live KV slot cheaply
(anthropic live continuation match=tool-output-ids). On the OpenAI
chat/completions endpoint the equivalent never fires: clients replay the
assistant tool-call turn re-serialised (thinking stripped to an empty
block, tool call re-encoded from JSON), so exact token/text prefix
matching fails at the previous generation's tail even though common
reaches the entire previous prompt. The live slot is discarded and
prefill restarts from an older disk checkpoint — O(context) per tool
round. Before (Open WebUI, native function calling):

live kv cache miss live=15139 prompt=20460 common=14903 reason=token-mismatch
kv cache stored tokens=15139 trimmed=0 reason=evict key=token-text size=221.61 MiB
chat ctx=5779..14903:9124 TOOLS prompt done 31.253s

The trailing role:"tool" message's tool_call_id is the same precise
continuation handle as Anthropic's tool_use_id, so this PR reuses that
fast path for API_OPENAI chat: collect trailing tool ids at parse time
(chat_prepare_live_continuation, mirroring the Anthropic prepare), and
when the ids and live token frontier match, keep the sampled KV and
append only EOS + tool results + next assistant prefix.

After, same client, same conversation shape, ~24k context:

chat live continuation match=tool-output-ids ids=1 cached=19597 prompt=22207
chat ctx=19597..22207:2610 TOOLS prompt done 11.234s
chat live continuation match=tool-output-ids ids=1 cached=23507 prompt=23552
chat ctx=23507..23552:45 TOOLS prompt done 0.798s
chat live continuation match=tool-output-ids ids=1 cached=24151 prompt=24164
chat ctx=24151..24164:13 TOOLS prompt done 0.230s

Tool rounds drop from 45-90s to 0.2-11s, and the per-round ~300 MiB
evict/store KV disk cycle disappears. A genuine mid-conversation prompt
change still evicts and re-prefills correctly (verified by editing
history mid-chat).

Design notes:

• State is stored in the existing anthropic_live_* request/slot
  fields; the matcher, remember, and clear sites are protocol-gated. A
  rename to something protocol-neutral would touch many more lines —
  happy to do that if preferred.
• The frontier is only remembered for chat when
  should_canonicalize_tool_checkpoint() is false, since
  canonicalization rewrites the live session and would strand the
  remembered frontier.
• Every gate failure returns 0 and falls through to the existing
  matching chain, so behaviour is unchanged whenever the fast path
  doesn't apply.
• Adds test_chat_live_tail_renders_tool_results_only covering the
  role:"tool" tail rendering and the protocol gate; full embedded suite
  passes.

Testing

• make test: full suite passes — all tracks including --server,
  --tool-call-quality, --long-context, and --logprob-vectors —
  on Apple M5 Max 128GB, Metal backend, macOS 26.5.2, model
  DeepSeek-V4-Flash-IQ2XXS-w2Q2K-AProjQ8-SExpQ8-OutQ8-chat-v2.gguf (q2).
• make cpu: builds clean (verified on x86-64 Linux).
• Live verification: same model fully resident on Metal, Open WebUI 0.10.x
  native function calling — eleven consecutive chat live continuation hits
  across chained tool rounds at 6k–24k context (0.2–11s per round vs 45–90s
  before), zero evict/store KV disk writes between rounds, and correct
  evict + re-prefill on a genuine mid-conversation prompt edit.