feat: Add declarative API for vector and full-text indexes

[Mini256]

close #89
close #103

Features

Declarative API:

• Vector Index is automatically created when using a VectorField.
• FullText Index is created automatically when using FullTextField.
• Supports passing VectorIndex / FullTextIndex objects in the __table_args__ field of the TableModel to define indexes.

Command API:

• Supports manual index creation by using VectorIndex("vec_idx", table.text_vec).create(db_engine)

Breaking changes

• Since the way to add a columnar replica is implemented using ADD_COLUMNAR_REPLICA_ON_DEMAND, it may not work properly with the OP version, which does not support the ADD_COLUMNAR_REPLICA_ON_DEMAND syntax.

Example code

Define vector index via VectorField (best practice)

class Chunk(TableModel):
    __tablename__ = "test_vector_index"
    id: int = Field(primary_key=True)
    text_vec: list[float] = VectorField(dimensions=3)    #  👈 index=True by default

tbl = client.create_table(schema=Chunk, mode="overwrite")
assert tbl.has_vector_index("text_vec")

Manual create vector index

class Chunk(TableModel):
    __tablename__ = "test_vector_index"
    id: int = Field(primary_key=True)
    text_vec: list[float] = VectorField(dimensions=3, index=False)

tbl = client.create_table(schema=Chunk, mode="overwrite")
tbl.create_vector_index(name="vec_idx_on_text_vec", column_name="text_vec")

Define vector index via FullTextField (best practice)

class Chunk(TableModel):
    __tablename__ = "test_fulltext_index"
    id: int = Field(primary_key=True)
    text: str = FullTextField()    #  👈 index=True by default

tbl = client.create_table(schema=Chunk, mode="overwrite")
assert tbl.has_fts_index("text")

[Icemap]

LGTM

[breezewish]

Two ideas.

1

It may be better to be:

self.client.db_engine.create_vector_index("vec_idx_on_text_vec", Chunk.text_vec)

2

Normally it is better to only have one way (or try best to have minimal ways) to achieve one thing, instead of several ways to achieve the same thing. Because the later one leads user to think, to choose, to learn difference, to have "best practice", which is a waste of time. (BTW this is the philosophy Golang follows). For this reason, possibly we only need two examples for now:

A. Define a vector field (with vector index) using VectorField
B. Define a vector field without vector index using VectorField

I do think it may be useful to create index lazily over an existing data. However we need to carefully think about under what scenario we can bring this feature to users. Needs some more work.

The rest LGTM.