fix(cli): enforce /v1/completions stop sequences host side for qairt

[Shishir Shrivastava (Shishir2405)]

Summary

The qairt plugin rejects any generate call that carries stop sequences (sdk/plugins/qairt/src/llm.cpp:219), but completion.go forwarded Stop unchanged to every runtime. Any FIM autocompletion client sending a stop list (which real editor clients always do) got ErrorCode[-100016] and an empty completion on QAIRT models, while the identical request worked on llama_cpp GGUF models.

This implements the host-side enforcement the issue proposed:

1. Stop is withheld from GenerationConfig when the resolved runtime is qairt.
2. A new stopMatcher (cli/server/handler/stop_match.go) matches the requested stop strings against generated text as it comes in, buffering only the minimal tail needed to catch a stop sequence split across token boundaries (e.g. <fim + _suffix + >).
3. On a match, generation is cancelled through the existing OnToken return-false path (already supported by the qairt pipeline, reported as stop_reason: "user" → finish_reason: "stop" via the existing mapFinishReason).
4. The returned text is truncated at the match, for both the streaming (SSE) and blocking response paths.

llama_cpp models are unaffected — Stop is still forwarded to the plugin exactly as before when the runtime isn't qairt.

Test plan

• Added stop_match_test.go: table-driven tests for stopMatcher, including the exact FIM stop-sequence-split-across-tokens scenario from the issue, multiple stop sequences, earliest-match-wins, empty stop strings, and the case where a partial match never completes (must still flush the held-back text).
• Verified the matcher logic in isolation (extracted to a throwaway module, no cgo/SDK dependency) — all cases pass, including the issue's repro shape.
• gofmt -l clean on all three changed/added files.
• Could not run go vet / go test ./cli/server/handler/... or bazelisk coverage in this environment — the handler package imports bindings/go, which requires the built SDK bridge (sdk/pkg-geniex/), and per _build-sdk.yml that only builds on windows-arm64 / linux-arm64 (Docker) / android-arm64, none of which were available here. I'd appreciate a CI run / local verification on a supported host before merge.

Notes

• Scope kept to /v1/completions, matching the issue. chat.go's /v1/chat/completions path doesn't forward Stop to GenerationConfig at all today (a separate gap, if it is one) — left untouched here to keep this PR focused.
• No new dependencies; stop_match.go only uses strings from the standard library, so go.mod/go.sum are untouched.

Closes #1341