feat(python): support DataFrame init from generators

[alexander-beedie]

Expands on #5411, adding support for initialising DataFrames from generators (in either orientation). Can even init from a generator of generators now, if that floats your boat ;)

from random import randint
import polars as pl

def gendata(n_rows, n_cols):
    for i in range(n_rows):
        yield (randint(-100,100) for j in range(n_cols))

pl.DataFrame( gendata(5,3), columns=["x","y","z"] )

# shape: (5, 3)
# ┌─────┬─────┬─────┐
# │ x   ┆ y   ┆ z   │
# │ --- ┆ --- ┆ --- │
# │ i64 ┆ i64 ┆ i64 │
# ╞═════╪═════╪═════╡
# │ 60  ┆ 26  ┆ -69 │
# │ 3   ┆ -58 ┆ 19  │
# │ -49 ┆ 76  ┆ 63  │
# │ 44  ┆ -93 ┆ -71 │
# │ 42  ┆ -2  ┆ 58  │
# └─────┴─────┴─────┘

It's a bit more complex than the previous Series update, as you don't necessarily know in advance how many columns you have; if that can't be inferred (eg: no 'columns' param) then the initial chunk is limited to 1000 rows - all subsequent updates can then target taking chunks of 1,000,000 elements, which initial testing shows seems to be reasonable, but we can probably optimise further based on the underlying/inferred schema.

Also:

• Calls to PyDataFrame.read_rows make better use of the relatively new schema override param, to avoid post-init casts.
• Fixed Series init from generators that return no data.
• Better generator type-detection.