fix: align /v1/chat/completions SSE stream with OpenAI spec

[sufubao]

Summary

• Add missing data: [DONE]\n\n terminator at the end of /v1/chat/completions streams — clients following the SSE spec were hanging waiting for this sentinel (the /v1/completions endpoint already emitted it, so this was an inconsistency within LightLLM itself)
• Emit a role-only initial chunk delta: {role: "assistant", content: ""} once per choice before any content, matching OpenAI API / vLLM behaviour
• Remove role: "assistant" from all subsequent delta chunks (plain content, tool-call text, and reasoning paths) so role appears only in the first chunk as the spec requires

Chunk sequence after this fix

data: {"delta": {"role": "assistant", "content": ""}, ...}   ← role-only first chunk
data: {"delta": {"content": "Hello"}, ...}                   ← token chunks (no role)
...
data: {"delta": {}, "finish_reason": "stop", ...}            ← finish chunk
data: {"choices": [], "usage": {...}, ...}                    ← only if include_usage=true
data: [DONE]

Test Plan

• Verify streaming response with a standard SSE client no longer hangs after generation completes
• Confirm first chunk carries role: "assistant" with content: ""
• Confirm subsequent content chunks do not repeat role
• Confirm data: [DONE] appears as the final line of the stream
• Test with stream_options: {include_usage: true} — usage chunk precedes [DONE]
• Test tool-call streaming — role-only chunk still emitted first, tool call chunks unaffected

[gemini-code-assist]

Code Review

This pull request updates the OpenAI API streaming implementation to comply with the SSE specification by emitting an initial role-only chunk and removing the role from subsequent deltas. It also adds a [DONE] terminator to the end of the stream. Feedback was provided to encode the terminator string to bytes to ensure consistency with the function's type hint and other parts of the codebase.