Fix TEI stale socket retries

[oldnicke]

Summary

Fixes #3284.

Remote TEI embedding requests can intermittently fail after idle periods when the pooled HTTP client reuses a socket that has died underneath it. This change makes TEI transport handling more resilient by:

• retrying retryable transport-layer OSErrors such as EBADF within the existing TEI retry budget,
• setting an explicit short keepalive expiry on TEI HTTP clients so idle pooled sockets are recycled sooner, and
• applying the shared TEI transport retry handling to both embeddings and the TEI reranker client.

Tests

• uv run pytest hindsight-api-slim/tests/test_tei_embeddings.py hindsight-api-slim/tests/test_tei_retry.py -q
• uv run pytest hindsight-api-slim/tests/test_tei_cross_encoder.py -q
• ./scripts/hooks/lint.sh

Notes

• No API schema changes.
• The regression test covers a first-request EBADF followed by a successful retry.