fix(utils): preserve embedding types across batches

[OllieinCanada]

Fixes #796

Summary

• collect embedding types from every batched response, not only the first
• continue skipping None values while preserving types returned by later batches
• add a regression for a later int8 response

Root cause

merge_embed_responses() built its field list from embeddings_type[0]. When that response did not contain a type that a later batch returned, the type was never added to the merged dictionary and its values were silently lost.

Validation

• regression fails on main (int8 is None)
• pytest -q tests/test_embed_utils.py (7 passed)
• mypy . (341 source files, no issues)
• git diff --check

The complete test collection requires CO_API_KEY in two existing modules, so no live API tests were run. Ruff reports the repository's existing import-format findings in both touched files; this patch does not reformat unrelated code.

---

Note

Low Risk
Small, targeted change to batch merge logic with a regression test; affects embed response assembly only, not API calls or auth.

Overview
Fixes batched embeddings-by-type merging so embedding types from later API batches are not dropped when the first batch did not return those types.

merge_embed_responses() now decides which fields to merge by checking whether any batch has a non-None value for each type, instead of only inspecting the first response. Merging still skips None per batch when concatenating vectors.

A regression test covers a case where the first batch has only float_ and a later batch adds int8, ensuring both types appear in the merged result.

^{Reviewed by Cursor Bugbot for commit 9806f64. Bugbot is set up for automated code reviews on this repo. Configure here.}