fix(memory-base): handle tagged Ollama embedding models (#14740)

fix(memory-base): strip Ollama tag suffix in embedding provider inference

Ollama's /api/tags returns model names with a tag (e.g. bge-m3:latest),
but the unified catalog stores them bare (bge-m3). The catalog lookup
and pattern fallback both failed to match tagged names, causing every
Ollama embedding model to be misclassified as OpenAI.

Strip the :tag suffix before catalog lookup and pattern matching so
that bge-m3:latest correctly resolves to 'Ollama'.

Fixes #14731

Co-authored-by: dymux <putramkti@users.noreply.github.com>