Implementations: pyspark and ray

[martindurant]

Remaining issues:

• pyspark picking up wrong python on windows (perhaps need environment variables set)
• time dtype inference requires new release of awkward, affects pyspark only
• ray partitioning isn't deterministic?

[martindurant]

(tests will pass following new release of awkward scikit-hep/awkward#3392 )

[martindurant]

Needs documentation updates. Refactoring to have the same code in all the lazy backends will be done in a follow-up (note how similar the spark and ray implementations are, that we wish to also do polars-lazy, and that the dask-dataframe version now appears less complete by comparison.

Also, we should add at least one numba test; should we have the same test suite for all backends?