refactor(python): move run-compression init into Client.__init__

[Angus Jelinek (angus-langchain)]

Summary

Compression setup previously ran inside tracing_control_thread_func after the first client.info fetch. This asynchronous gate meant runs created before the server-flag handshake completed lived in tracing_queue, while later runs went to compressed_traces — a race that required a defensive fix in #2796.

The server-side zstd_compression_enabled flag is now assumed to be set on every supported deployment (cloud and all self-hosted versions), so the asynchronous gate is no longer needed.

Changes

• Compression state (compressed_traces, _data_available_event, _futures) is initialized in Client.__init__ using only local signals: ZSTD_AVAILABLE, DISABLE_RUN_COMPRESSION, tracing mode, auto_batch_tracing, and user-provided batch_ingest_config.use_multipart_endpoint.
• Both the tracing and compression background threads spawn from __init__, so compressed_traces is either set or None for the lifetime of the client — no mid-flight transition.
• The server-flag check and the "Please update to the latest version of LangSmith" warning are removed from _background_thread.py. num_known_refs in the tracing thread is now computed from client.compressed_traces is not None.
• tracing_queue still exists and handles per-request auth/URL overrides, OTel mode, and the ZSTD-not-installed fallback. The race fix from fix(python): flush both tracing_queue and compressed_traces in flush() #2796 stands as defense in depth.

Test changes

• parse_request_data (tests/unit_tests/conftest.py) now detects zstd frames and streams-decompresses them, so existing payload assertions keep working without per-test changes.
• A few tests that assert the non-compressed multipart payload structure (attachments, raw bytes data) now opt out via LANGSMITH_DISABLE_RUN_COMPRESSION=true rather than relying on the (now-gone) server-flag behaviour.
• test_create_run_without_compression_support is deleted — the server-doesn't-support-compression scenario no longer exists.
• test_client_gc uses client.flush() instead of client.tracing_queue.join() so it drains both queues.