Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: usamoi <usamoi@outlook.com>

Consider the following scenario:

1. Process 1 performs a search and finds vector A. It locates `x` (1920
dimensions).
2. Process 2 deletes vector A from the index. `x` (1920 dimensions) and
`y` (1 dimension) are deleted.
3. Process 3 inserts vector B into the index. `y` (1920 dimensions) and
`z` (1 dimension) are inserted.
4. Process 1 locates `y` (1920 dimensions) and detects a length
mismatch, resulting in an error.

Since vector a can be deleted and is therefore guaranteed not to be
included in the query results, any error related to a during reranking
should be silently ignored.

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: xieydd <xieydd@gmail.com>

cc @xieydd

---------

Signed-off-by: Keming <kemingyang@tensorchord.ai>

adds a GUC `io_rerank`, with the following possible values:

* `read_buffer`: indicates a preference for `ReadBuffer`.
* `prefetch_buffer`: indicates a preference for both `PrefetchBuffer`
and `ReadBuffer`; it's good for disk vector search; this is default on
PostgreSQL 13, 14, 15, 16.
* `read_stream`: indicates a preference for `read_stream`; it's good for
disk vector search; this option is only available in PostgreSQL 17; this
is default on PostgreSQL 17.

notes:

* prefetching distance of `prefetch_buffer` depends on the compile-time
constant `WINDOW_SIZE`.
* prefetching distance of `read_stream` depends on the GUC parameter
`effective_io_concurrency`.
* prefetching for the bit vector has not been implemented.
* prefetching for the centroid has been implemented, but disabled.

Signed-off-by: usamoi <usamoi@outlook.com>

AVX512 (probably excluding AVX512FP16) is being stabilized. With the
next compiler version bump, we might finally be able to get rid of
nightly toolchain.

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: usamoi <usamoi@outlook.com>

Adds a GUC `io_search`, with the same possible values of `io_rerank`.

notes:

* prefetching for the bottom-level bit vectors is implemented.
* prefetching for the other levels of bit vectors is not implemented
yet.

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: xieydd <xieydd@gmail.com>

## Feat
- Enable `rerank_in_table=true` with `read_stream`(pg17) and
`prefetch_buffer`(not pg17)

## Optimization
- Remove redundant matrix multiply in both heap query/insert,
accerate most searches

## Chore
- Rename `vchordrq.prererank_filtering` into `vchordrq.prefilter`

## Test
- Add specialized tests for pg16(prefetch_buffer) and pg17(read_stream)

---------

Signed-off-by: cutecutecat <junyuchen@tensorchord.ai>

note:
- residual quantization for inner product is implemented

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: usamoi <usamoi@outlook.com>

Signed-off-by: usamoi <usamoi@outlook.com>