MEAI adapter: ToUsageDetails drops Usage.OutputTokensDetails.ThinkingTokens (UsageDetails.ReasoningTokenCount never populated)

The Microsoft.Extensions.AI adapter (AnthropicClientExtensions.AnthropicChatClient, verified on 12.40.0) never forwards Usage.OutputTokensDetails.ThinkingTokens. The response model carries the count (final Usage and streaming MessageDeltaUsage both), but ToUsageDetails forwards only input/output/cache-creation/cache-read and the two server-tool counters. So a consumer on the MEAI abstraction cannot meter reasoning tokens (for example, the OTel GenAI gen_ai.usage.reasoning.output_tokens attribute, added in semconv v1.41.0, is unreachable from this surface).

UsageDetails has had a first-class home for this since M.E.AI.Abstractions 10.1.x: ReasoningTokenCount (dotnet/extensions#7122), documented as counted within OutputTokenCount; the same semantics as the Anthropic field (thinking_tokens is documented as always ≤ output_tokens, with output_tokens remaining the inclusive billing total). The peer OpenAI adapter already populates it (OpenAIChatClient.FromOpenAIUsage: ReasoningTokenCount = tokenUsage.OutputTokenDetails?.ReasoningTokenCount), and the 12.40.0 package already depends on M.E.AI.Abstractions >= 10.5.1, so the fix looks like a one-liner: ReasoningTokenCount = usage.OutputTokensDetails?.ThinkingTokens plus the streaming-delta merge path.

Context: #221 is adjacent request-side reasoning evidence (the adapter's ChatOptions.Reasoning mapping).

Happy to PR this. Our downstream currently records the count as unreachable, with a pinned boundary test that alerts us if/when a release fixes it.