Feature/instance udfs

[timsaucer]

Which issue does this PR close?

Closes#822

Rationale for this change

With this change we add an additional argument to udaf and udwf to be used for calls to __init__ of the class that implements the aggregation or partition evaluator. This allows users to pass in these arguments so we can reuse a single class that requires parameters.

The other option is to always pass in these parameters as lit() values which is not as performant.

What changes are included in this PR?

Updates the udaf call to add in the provided arguments when initializing the classes. Also adds unit tests.

Are there any user-facing changes?

There is a non-breaking change to the udaf function that allows for an optional set of arguments to be passed in. If any users are instead calling AggregateUDF() directly they will need to add in an empty list as the extra argument.

There is a breaking change to udwf but since it is not in any released version yet, it shouldn't count as a user-facing change in my opinion.

[Michael-J-Ward]

Instead of saving the __init__ arguments and calling them ourselves, wouldn't it be more flexible and less work for datafusion-python to allow the user to provide a factory function?

Modulo runtime assertions, this works on main currently:

deftest_udaf_aggregate_with_arguments(df):
bias=10.0deffactory() ->Accumulator:
returnSummarize(bias)
summarize=udaf(
factory,
pa.float64(),
pa.float64(),
[pa.float64()],
volatility="immutable",
# arguments=[bias],
)
df1=df.aggregate([], [summarize(column("a"))])
# execute and collect the first (and only) batchresult=df1.collect()[0]
assertresult.column(0) ==pa.array([bias+1.0+2.0+3.0])

[timsaucer]

This is a very good idea. I'll change to your approach.

[Michael-J-Ward]

Looks great!

[Michael-J-Ward]

Ah, shoot. Sorry @timsaucer, I created a conflict when I merged the release-testing PR. This needs a rebase.

[timsaucer]

No worries at all. I'll wait for #892 CI to finish, merge that, then rebase this.