fix(stream): parse streamed tool-call chunks (fixes 'dict' object has no attribute 'delta')

[KillerQueen-Z]

Problem

Streaming a tool call crashes the SDK:

File ".../blockrun_llm/solana_client.py", in _aiter_and_archive
    if choice.delta.content:
AttributeError: 'dict' object has no attribute 'delta'

Plain chat streams fine; any tool call breaks. Hits LLMClient and SolanaLLMClient, sync and async. Surfaced via blockrun-litellm's sidecar but the bug is here in the SDK.

Root cause

ChatChunkDelta.tool_calls used the strict non-stream ToolCall (id / function.name / arguments all required). But OpenAI streams tool calls incrementally:

• first frame: index + id + function.name (+ empty args)
• later frames: only index + function.arguments fragments (no id, no name)

Those argument-fragment frames fail ToolCall validation → ChatCompletionChunk(**chunk) raises → _iter_sse_chunks falls back to ChatCompletionChunk.model_construct(...), which doesn't parse nested models, so choices stay as raw dicts → the archive loop's choice.delta.content blows up.

Reproduced at the type level: a {"tool_calls":[{"index":0,"function":{"arguments":"{\"ci"}}]} frame raised ValidationError, fell back to model_construct, and choice.delta then raised AttributeError.

Fix

1. Lenient streaming tool-call types — add ChatChunkFunctionCall / ChatChunkToolCall (all fields optional + index) and use them in ChatChunkDelta.tool_calls. Streamed tool-call frames now parse into real objects instead of triggering the model_construct fallback. The non-stream ToolCall (used by ChatMessage) stays strict.
2. Harden the archive loops (defense in depth) — the four _iter_and_archive / _aiter_and_archive loops (client.py + solana_client.py, sync + async) now use dict-tolerant accessors (stream_choice_content / stream_choice_finish_reason / chunk_usage_dict), so a model_construct fallback for any future reason can no longer crash the stream.

Tests

New regression tests (sync + async) feed argument-fragment frames through the paid streaming path (so cost_usd > 0 and the archive loop actually runs). They crash on the old code (AttributeError: 'dict' object has no attribute 'delta') and pass with the fix; arguments reassemble to {"city":"Paris"} and finish_reason=tool_calls.

• Full unit suite: 253 passed.