fix(tracing): sanitize non-finite span payloads

[sylvesterkaczmarek]

This pull request fixes OpenAI tracing normalization so non-finite float values cannot bypass the existing JSON sanitizer merely because the surrounding input or output field is small.

BackendSpanExporter._value_json_size_bytes() now measures payloads using strict JSON semantics with allow_nan=False. A field containing NaN, Infinity, or -Infinity therefore follows the existing unserializable-value path, where _sanitize_json_compatible_value() removes the invalid nested value before the tracing payload is sent. Finite JSON payloads keep the existing fast path and are unchanged.

The change deliberately reuses the current OpenAI-ingest sanitization boundary instead of introducing another recursive traversal. Focused regression coverage exercises both input and output, all three non-finite float forms, preservation of valid sibling data, and a finite control case.

This pull request resolves #4551.

[seratch]

Closing this for the reason at #4551 (comment)