feat(qairt): native stop sequences in the geniex-qairt pipeline

[MAN$I VERMA (mansiverma897993)]

Fixes #1341

FIM clients always send a stop list, /v1/completions forwarded it to the plugin unchanged, and the qairt plugin rejects any generate call with stop sequences (-100016), so completions failed on every QAIRT model.

For non-llama_cpp runtimes the stop list is no longer forwarded. A stopScanner enforces it host-side: it matches the stop strings against the streamed tokens, holds back the longest tail that could still start a match so a stop spanning token boundaries never leaks, cancels generation through the token callback on a match (stop_reason: "user", mapped to finish_reason: "stop"), and truncates the output at the match. Applied to both the streaming and blocking paths. llama_cpp keeps its native stop handling, so existing behavior is unchanged there.

Test report

TestStopScanner (added in completion_test.go), 7 cases: no match, match inside one token, match spanning tokens, earliest of multiple stops wins, stop at start, empty stop strings pass through, held tail flushed at end.

Verified end-to-end without the SDK toolchain via a pure-Go stub harness (exact-signature geniex_sdk stubs, real gin HTTP server, fake plugin that emits tokens through OnToken and honors cancellation), same pattern as #1317:

--- PASS: TestE2EQairtStreamingStopSpanningTokens   stop split across tokens; SSE text truncated, finish_reason stop, plugin got Stop=nil, generation cancelled early
--- PASS: TestE2EQairtStreamingNoMatchFlushesTail   held-back tail emitted when generation ends without a match
--- PASS: TestE2EQairtStreamingLength               finish_reason length preserved when max_tokens is hit
--- PASS: TestE2EQairtBlockingStop                  blocking JSON response truncated at the stop, finish_reason stop
--- PASS: TestE2EQairtBlockingNoStopsUnchanged      no stops: no OnToken registered, response unchanged
--- PASS: TestE2EQairtIssue1341Replay              exact request from #1341 (10-entry FIM stop list, streaming): 200, truncated completion, no -100016
--- PASS: TestE2ELlamaCppStopsForwardedToPlugin     llama_cpp still receives the stop list natively
--- PASS: TestCompletionPrompt / TestCompletionStop / TestStopScanner / TestCompletionUnsupported
ok      github.com/qualcomm/GenieX/cli/server/handler

go vet clean, gofmt clean.

[RemiliaForever (RemiliaForever)]

The clean fix is native stop support in the geniex-qairt pipeline — generateTokens in llm_pipeline.cpp already decodes per-token, accumulates full_text, and reports stop_reason, so it'd match llama_cpp and cover every consumer. In the plugin or handler it's just a byte-level patch outside the decode loop.

[MAN$I VERMA (mansiverma897993)]

Thanks for the review RemiliaForever (@RemiliaForever)! Addressed - stop is now handled natively in the geniex-qairt pipeline instead of being enforced host-side.

What changed:

• generateTokens in llm_pipeline.cpp now matches the stop sequences byte-wise against the streamed output. A StopMatcher holds back the longest tail that could still start a match (so a stop spanning token boundaries never leaks through the token callback), cancels generation on a match, truncates the output at the match, and reports stop_reason: "stop_sequence" - matching llama_cpp and covering every consumer of the pipeline.
• The qairt plugin adapter (sdk/plugins/qairt/src/llm.cpp) now forwards the FFI stop list into GenerationConfig::stop_sequences instead of rejecting it with PARAM_NOT_SUPPORTED (-100016), and maps the new stop_sequence stop reason through to profile_data.
• The temporary host-side stopScanner in the /v1/completions handler is removed; the handler forwards the stop list to every plugin as it did before the workaround. llama_cpp keeps its native handling unchanged.
• Bumped third-party/geniex-qairt to the plugin commit with native stop support (mansiverma897993/geniex-qairt-plugin feat/native-stop), which includes the StopMatcher unit tests (byte-level matching, boundary-spanning stops, earliest-stop-wins, held-back tail flush).

Fixes #1341 either way - no more -100016 on QAIRT models.