Seperate quantizer training from index build

[syvb]

This splits out quantizer training from do_heap_scan. This doesn't change any of the behaviour of index-building, but is going to be useful in a future PR that adds support for parallel indexing.

This PR also makes it so the creation progress is set properly when building a plain (non-quantized) index.

[Copilot]

Pull Request Overview

This PR separates the quantizer training phase from the main heap scan in the index build process and ensures the index-creation progress is updated correctly for plain storage.

• Removed start_training/add_sample/finish_training methods from the Storage trait and its implementations.
• Introduced a standalone train_quantizer function (with BqTrainState) to handle quantizer training before graph building.
• Updated ambuild and do_heap_scan to accept and forward WriteStats, reflecting the split between training and building.

Reviewed Changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/access_method/storage.rs	Dropped unused training methods from the Storage trait.
src/access_method/sbq/storage.rs	Switched SBQ storage to load a pretrained quantizer; removed inline training methods.
src/access_method/plain/storage.rs	Removed no-op training stubs to align with trait cleanup.
src/access_method/graph/mod.rs	Deleted get_meta_page_mut accessor as it’s no longer used.
src/access_method/build.rs	Added train_quantizer, restructured ambuild for two-phase building, and updated do_heap_scan.

Comments suppressed due to low confidence (3)

pgvectorscale/src/access_method/build.rs:32

• [nitpick] The struct name BqTrainState is inconsistent with the Sbq prefix used elsewhere; consider renaming it to SbqTrainState for clarity and consistency.

struct BqTrainState<'a, 'b> {

pgvectorscale/src/access_method/build.rs:277

• [nitpick] Consider adding a doc comment to train_quantizer to explain its purpose, parameters, return value, and when it should be used.

fn train_quantizer(

pgvectorscale/src/access_method/build.rs:277

• It would be beneficial to add unit or integration tests covering train_quantizer to ensure the quantizer training logic and metadata storage behave as expected.

fn train_quantizer(

[tjgreen42]

Looks good to me, modulo minor comment about naming!