Normalize finish_reason across providers

[cpsievert]

Mirrors tidyverse/ellmer#1015.

What's now possible

turn.finish_reason returns a consistent vocabulary regardless of which provider you're using:

Value	Meaning
"success"	Normal completion
"max_tokens"	Hit token limit
"content_filter"	Blocked by safety filter
"stop_sequence"	Hit a stop sequence
"context_window"	Context window exceeded

Unknown provider-specific reasons pass through as-is so callers can still inspect them.

Changes

• Each provider module gains a normalize_finish_reason() function with a mapping dict keyed to the upstream API docs
• AssistantTurn.finish_reason is now set by calling that function rather than storing the raw provider string
• OpenAI streaming fix: response.incomplete events are now handled in stream_merge_chunks so truncated streaming responses are returned rather than silently dropped
• All finish_reason == "stop" / "end_turn" / "STOP" assertions in provider tests updated to "success"
• Unit tests added for each provider's normalization function covering known mappings, unknown pass-through, and None handling

[thisisnic]

FYI, I've extended the list of standard values in tidyverse/ellmer#1039 to also include "tool_use" as it was raising unnecessary warnings.