feature: Document tensor type migration workflow: float32 → bfloat16 without re-feeding

[wanliAlex]

Following the merge of #35674 (support tensor type conversion from float to bfloat16), we'd like to have a feature for migrating existing tensor<float> fields to tensor<bfloat16> in production without re-feeding all documents.

Use case

Users who have an existing tensor<float>(x[N]) field and want to reduce storage/memory by converting it to tensor<bfloat16>(x[N]), without re-feeding the entire corpus.

Approaches discussed

Option A – Point queries to the new field name:

1. Create field-b (bfloat16) with field-a as input
2. Reindex to populate field-b
3. Update queries/ranking profiles to reference field-b

Option B– Refeed:

Update the schema and re-feed the documents.

Request

It would be nice to support in-place conversion from float -> bfloat16 without refeeding or query pointing.