Sketch out quantization tasks and calls into DiskANN

[kevin-montrose]

Change expected return from DiskANN's insert so we can signal a request to quantize.

Spin up tasks to pump background quantization when that is requested.

Tasks take two forms:

• A single task per Vector Set index which calls build_quant_table(context, index)
• After that finishes and true is returned, some number that proceed in parallel and call backfill_quant_vectors(context, index, task_index, task_count)

It's on DiskANN to find vectors to backfill and split them up based on task_index / task_count.

Since locks are not held over the entire backfill, DiskANN also needs to validate that quantization is still needed when build_quant_table is called (recreation and restarts can mess with that) and backfill_quant_vectors needs to gracefully deal with concurrent calls.

---

Upon further discussion, we're also cleaning up quantizer options and the FFI between Garnet and DiskANN around vector formats.

Namely:

• Rename quantizer XPREQ8 -> XNOQUANT_U8
• Add XBIN_I8
• Rename format XB8 -> XU8
• Add format XI8
• Garnet now knows the "natural" format of different quantizers and only passes vectors of those formats to DiskANN in insert & search calls

Old options are aliased so we can keep XPREQ8 and XB8 in tests and whatnot without issue.