feat(sdk): return PROMPT_TOO_LONG when the prompt itself overflows

[RemiliaForever (RemiliaForever)]

Summary

Overlong input previously always mapped to GENIEX_ERROR_LLM_TOKENIZATION_CONTEXT_LENGTH, so callers could not distinguish a prompt that never fit the context window from a window that filled up mid-generation. This PR splits the two:

• Prompt too long — the prompt alone exceeds the context window (even after a context shift, or the model can't shift). Now returns the new GENIEX_ERROR_LLM_GENERATION_PROMPT_TOO_LONG from both the llama_cpp and qairt LLM/VLM plugins.
• Context length exceeded — the in-flight generation fills the window mid-decode. Keeps returning CONTEXT_LENGTH.

Downstream behavior:

• SDK plugins — llama_cpp reports the phase-appropriate error from its shared process() (prefill → prompt-too-long, decode → context-length); qairt reads the pipeline's new "prompt_too_long" stop_reason.
• Go binding — exposes ErrLlmGenerationPromptTooLong.
• Server — prompt-too-long is a 400 client error with no partial output; a window exhausted mid-generation is a normal truncated completion (finish_reason=length, HTTP 200) on both blocking and streaming paths, for chat and completion.
• CLI — a distinct hint and a reset-and-continue path for prompt-too-long.

Sliding-window support is unchanged: the error is reported only when prefill actually fails, never via a pre-check.

The qairt plugin behavior depends on the companion submodule change (qualcomm/geniex-qairt-plugin#34); this PR bumps the third-party/geniex-qairt pointer to include it.

Test plan

• llama_cpp end-to-end (local): prompt-too-long, sliding-window intact, server 400 for prompt-too-long, server 200 finish_reason=length for mid-generation fill (blocking + streaming)
• qairt runtime (qcs, Qwen3-1.7B): CLI prompt-too-long vs context-length messages, --sliding-window continues, server 400, server 200 finish_reason=length
• qairt unit tests (qcs): ctest -L unit — ThrowsWhenPromptExceedsContext, SlidingWindowDisabledByDefaultStillThrows, new ThrowsWhenGenerationExceedsContext, 8/8 pass
• CLI process_test.go: prompt-too-long resets and continues
• CI: lint → build-sdk → CLI/Python/Android → docker
• Depends on feat(core): distinguish prompt-too-long from mid-generation overflow geniex-qairt-plugin#34 merging first (submodule pointer)

Depends on qualcomm/geniex-qairt-plugin#34