feat: duckdb support

[machow]

re-implementation of #380

Notes:

• This requires the code on master of https://github.com/Mause/duckdb_engine.
• Contains a patch to fix the issue in feat(sql): implement basic duckdb dialect #380 of literal ints being interpreted as strings. Basically, for duckdb we bind literals when executing the query, so that there are no parameters. Once Integer parameters sometimes become varchar when using python dbapi duckdb/duckdb#2972 is released, can remove the patch.

Edit: Updated notes (May 30th):

• Breaking: LazyTbl now gets the dialect name from sqlalchemy's URL, to work around needing dialects to set the correct name. This means LazyTbl must take an Engine, not a connection (which seems weird to pass it anyway).
• I added a temporary _is_dialect_duckdb function to handle these edge cases:
  • collect() uses duckdb's query(...).to_df() method
  • collect() binds literals when compiling query
  • assert_equal_query coerces a duckdb to_df() result to use 64 bit column types (e.g. int64, float64). This is to match pandas to_sql behavior (which coerces everything to 64 bit). This is just to make testing easier. Could annotate duckdb translations with result_type="variable" to tell the tests not to check exact dtypes.

Example

from siuba.data import mtcars
from siuba.sql import LazyTbl
from siuba import _, mutate, group_by, filter

from sqlalchemy import create_engine

engine = create_engine("duckdb:///:memory")
engine.execute("register", ("some_df", mtcars))

tbl = LazyTbl(engine, "some_df")

# mutation ----
tbl >> mutate(res = _.hp - _.hp.mean(), res2 = _.res * 2)

# group by cylinder, and filter all entries less than the mean ----
tbl >> group_by(_.cyl) >> filter(_.hp < _.hp.mean())

[machow]

AFAICT duckdb is quite a bit faster! In this filter example below..

edit: wait--nevermind, I need to dig a bit deeper into timings

[machow]

rebased on main

[machow]

Alright--should be working! I made a copy of the duckdb timings colab notebook, and added siuba code.

I'm super interested in shifting a lot of focus to supporting duckdb as a backend, and am super curious to explore nesting/unnesting with it!