fix(anthropic): capture extended-thinking reasoning tokens

[paultancre-bt]

Problem

The direct Anthropic integration (integrations/anthropic/) never records reasoning tokens. extract_anthropic_usage (_utils.py) only reads flat, top-level usage fields:

_ANTHROPIC_USAGE_METRIC_FIELDS = (
    ("input_tokens", "prompt_tokens"),
    ("output_tokens", "completion_tokens"),
    ("cache_read_input_tokens", "prompt_cached_tokens"),
    ("cache_creation_input_tokens", "prompt_cache_creation_tokens"),
)

But Anthropic reports extended-thinking usage nested, under usage.output_tokens_details.thinking_tokens (a subset of output_tokens). So completion_reasoning_tokens was always absent on thinking-model spans.

Flagged by the same customer as the sibling pydantic-ai bug ("the direct Anthropic integration records no reasoning tokens at all"). This is the follow-up called out in #745 / SDK-328.

Fix

Extract the nested output_tokens_details.thinking_tokens as the normalized completion_reasoning_tokens metric, matching the OpenAI/pydantic-ai integrations.

Verification

• Cassette-backed regression test recorded against a real Anthropic thinking response (claude-haiku-4-5, thinking enabled), usage came back with output_tokens_details.thinking_tokens: 168, and the span metric now matches.
• Red→green confirmed: without the fix the test raises KeyError: 'completion_reasoning_tokens'.
• Existing usage/metrics/cache tests still pass (7/7).

Blast radius

• Additive: one new metric read; no existing field/metric changes.
• Gated on include_output, so the streaming input-only usage path (include_output=False) is unaffected — verified it stays absent there.
• Degrades cleanly when output_tokens_details is absent: non-thinking requests don't return it, and older SDKs (e.g. anthropic 0.48.0) have no such field → skipped. Test is version-gated to latest, same pattern as the existing prompt-cache-TTL tests (no 0.48.0 cassette needed).

Relates to #745