perf: cache SDK version for span origin context

[Nathan (nathan-mialabs)]

cache installed Braintrust SDK version for the lifetime of the process
add regression test proving repeated span-origin merges resolve package metadata only once

why:

merge_span_origin_context() runs for every span, but currently resolves the installed Braintrust version through importlib.metadata each time. That triggers synchronous os.stat() calls. We observed this with AIOCOP, an asyncio event-loop monitoring tool: these synchronous calls appeared in stalls lasting 2.45–3.18 seconds under load.

Since the installed version cannot change during a process's lifetime, caching it removes repeated blocking I/O while preserving behavior.

Example stack trace captured in SigNoz:

importlib_metadata/__init__.py:838:mtime
<- importlib_metadata/__init__.py:833:search
<- importlib_metadata/__init__.py:974:<genexpr>
<- importlib.metadata:397:from_name
<- importlib.metadata:862:distribution
<- importlib.metadata:889:version
<- braintrust/span_origin.py:103:_sdk_version
<- braintrust/span_origin.py:92:merge_span_origin_context
<- braintrust/logger.py:4619:__init__
<- braintrust/logger.py:4761:start_span
<- braintrust/logger.py:2981:start_span

timing:

Local timing on macOS with Python 3.12 using python -m timeit:

• uncached importlib.metadata.version("braintrust"): 113 µs per call
• cached _sdk_version(): 10.3 ns per call
• approximately 11,000x lower hot-path overhead after the first lookup