Refactor index operations to be fully transactional

Currently, if there are multiple indexes of the same name, we handle index deduplication at https://github.com/lancedb/lance/blob/2280e0fbf9cf83a717dea08a97eb9ab62a6a07c9/rust/lance/src/dataset/transaction.rs#L1488-L1495.

This works for now because index creation and reindexing (CreateIndex(replace=true)) is typically done through a single recurring process. However, if we want to be strict, this would break if there are 2 concurrent reindex process that both modifies the same index. In that case, the second process will clobber the result of the first one, whereas the right behavior should be that the second one fail because of modifying the same index.

The right way is probably to pass the replaced index to the transaction operation at https://github.com/lancedb/lance/blob/main/rust/lance/src/index.rs#L415, and then update the conflict_resolver to properly resolve such conflict.

Read rust/lance/src/dataset/transaction.rs around the index deduplication logic and rust/lance/src/index.rs around line 415. Trace how the replaced index reaches the transaction operation and how conflict_resolver handles it; done means concurrent reindex operations on the same index conflict instead of one clobbering the other.