fix(server): unify generation failure responses

[Graffioh]

Summary

• add a server-owned ResponseError boundary that maps typed backend failures once
• stop classic generation failures before success telemetry, cache memory, SSE completion, or HTTP 200 response construction
• emit native OpenAI, Anthropic, and Responses API failure shapes for streaming and non-streaming clients
• replace the scheduler's separate failure flag and error string with one optional error value
• make success and failure terminals mutually exclusive in SseEmitter

Design

This keeps the current execution boundaries intact:

• ModelBackend::generate() remains the classic whole-request interface
• SeqEngine::step() remains the concurrent iteration interface
• GGML and backend code do not receive HTTP, JSON, or SSE types
• socket ownership stays in the classic worker and scheduler

The PR deliberately keeps ParsedRequest, GenerateRequest, GenerateResult, ServerJob, SchedSlot, and DaemonIO. It introduces a name only for the missing server-side failure consequence instead of renaming the pipeline or adding a common generation hierarchy.

This is complementary to #688 and does not touch its backend-configuration files.

Client behavior

• OpenAI-compatible streams receive an error object followed by [DONE]
• Anthropic streams receive an error event and no success message_stop
• Responses streams receive response.failed and no response.completed
• non-streaming failures use a mapped 400, 500, or 503 status with the matching API error envelope

Verification

• C++17 syntax checks pass for response_error.cpp, sse_emitter.cpp, http_server.cpp, and scheduler.cpp
• a model-free executable built from the production response-error and SSE sources passes all three protocol failure cases and terminal exclusivity
• git diff --check passes
• full GPU-linked test_server_unit is left to CI because the local host has neither CUDA nor HIP

[cubic-dev-ai]

All reported issues were addressed across 9 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 9 files

<sub>Requires human review: Auto-approval blocked by 3 unresolved issues from a previous review of this commit.

Re-trigger cubic</sub>