port the numba integration prototype from Enterprise to OSS, fixes feature-523

[jmao-denver]

this prototype in its current design must be used in an unpublished, explicit way. So it should not break anything visible to users.

[devinrsmith]

Looks good, mostly. Couple changes needed. Hard for me to tell if there are any big departures from how I originally structured my patch?

I'll be testing this out locally now.

[devinrsmith]

Great - was able to test the following, and it works:

from deephaven import TableTools, java, numba as dnb

@dnb.vectorize
def my_manual(x: java.int, y: java.int) -> java.int:
  return (x % 3) + y

IJ = TableTools.emptyTable(1000000) \
    .view("I=(int)i", "J=(int)(i * 2)") \
    .update(java.formula(my_manual, name = "K", column_names = ["I", "J"]))

[devinrsmith]

Good example of the chunking working:

from deephaven import TableTools, java, numba as dnb

@dnb.vectorize
def my_vectorized(x: java.int, y: java.int) -> java.int:
  return (x % 3) + y

vectorized_sum = TableTools.emptyTable(100000000) \
    .view("I=(int)i", "J=(int)(i * 2)") \
    .view(java.formula(my_manual, name = "K", column_names = ["I", "J"])) \
    .sumBy()

is very fast

vs

from deephaven import TableTools, java

def my_plain(x: java.int, y: java.int) -> java.int:
  return (x % 3) + y
  
plain_sum = TableTools.emptyTable(100000000) \
    .view("I=(int)i", "J=(int)(i * 2)") \
    .view("K = (int)my_plain(I, J)") \
    .sumBy()

is very slow.