fix: resolve namespaced auto model identifiers

[tandede]

Summary

• resolve exact namespaced Hub model IDs before falling back to directory names
• support local model and checkpoint directories through unambiguous basename matching
• normalize trailing path separators consistently for auto embedders and rerankers

Problem

AUTO_RERANKER_MAPPING contains four namespaced model IDs, but FlagAutoReranker.from_finetuned() previously stripped the namespace before looking them up. As a result, every one of those supported entries raised ValueError before model loading began. The same basename-only parsing also failed for local paths whose mapped key includes a namespace, while trailing separators produced an empty model name.

The resolver now prefers an exact mapping key, preserves checkpoint parent resolution, and only uses a namespaced basename alias when it is unique. Ambiguous basenames remain unresolved instead of selecting an arbitrary model class.

Test plan

• pytest -q tests (14 passed, 1 skipped with transformers 4.57.6)
• ruff check --select E4,E7,E9,F,I on changed Python files
• ruff format --check on the new resolver and regression test
• python -m compileall -q FlagEmbedding/inference tests/test_auto_model_resolution.py