add btx implementation to execute span spec tests

[Andrew Kent (realark)]

Needed this to test gzip compression and attachment processors. Breaking out into a separate PR for ease of review.

Adds btx test execution framework.

• pinned to latest spec: https://github.com/braintrustdata/braintrust-spec/tree/v0.0.7
• supports live end-to-end testing
  • VCR_OFF=true bundle exec rake test:btx:ci
  • https://www.braintrust.dev/app/Braintrust%20SDKs/p/ruby-unit-test/logs?range={%22from%22:%222026-06-01T20%253A17%253A03.000Z%22,%22to%22:%222026-06-01T20%253A22%253A03.000Z%22}
• also supports vcr-mode for replay (runs as part of the standard ci job)

PRODUCT CHANGES MADE

NEW FEATURES

• When Anthropic returns the per-TTL cache_creation breakdown, the SDK now emits prompt_cache_creation_5m_tokens / prompt_cache_creation_1h_tokens and drops the aggregate prompt_cache_creation_tokens (to avoid double-counting). Previously only the aggregate was emitted.
• tag openai spans as llm spans

MINOR BREAKING CHANGES

To bring the ruby sdk in line with the specs, a few changes to the anthropic tags were made

• output is now a single object {role, content} instead of a one-element array [{role, content}]
• The system prompt is now appended last ([user…, system]) instead of first ([system, user…]).
• Streaming now produces a span named anthropic.messages.stream instead of anthropic.messages.create