fix: offload SDMX response parsing from the event loop

[ypldan]

Applicable issues

• fixes Offload SDMX response parsing from the event loop #500

Description of changes

SDMX responses were parsed synchronously on the event loop; large data messages take seconds of CPU-bound parsing and stall every concurrent request. Both parse paths now run in a worker thread via asyncio.to_thread, mirroring the existing _data_msg_to_dataframe offload in dataset.py. Threaded XML parses in AsyncSdmxClient all run with structure=None and build private structures, so they stay fully parallel; the proxy v30 data parse carries a shared DataStructureDefinition that the reader attaches to the message and mutates while parsing (msg.structure...getdefault(...)), so dsd-bearing parses are serialized with a module-level threading.Lock.

• Run _parse_response via asyncio.to_thread in AsyncSdmxClient._fetch
• Run the StatGptSdmxProxyDataReader convert via asyncio.to_thread in AsyncStatGptSdmxProxyClient._proxy_data
• Add _PARSE_DSD_LOCK guarding reader.convert when a dsd is passed (no _parse_response caller currently does; the proxy data path does)
• Add unit tests asserting a slow in-thread parse does not block a concurrent coroutine (XML client and proxy client)

Checklist

• Title of the pull request follows Conventional Commits specification

By submitting this pull request, I confirm that my contribution is made under the terms of the MIT license.