CU-8696nbm03: Remove unigram table

[mart-r]

The unigram table we've been using is massive (10M-100M element array).
As such, when this gets saved to disk, it takes up a lot of space.

But really, all it is is a long array of numbers that are repeated based on the frequency they're expected (across all words).
E.g if you had words with equal counts, the number of their indices would be the same in the unigram table. And if you had some that had far higher counts, they would be far frequent in the unigram table.

What this PR does is remove the unigram table in favour of another approach.
This new approach finds the frequencies of all words, and then finds the cumulative probabilities (which is a sorted array starting near 0 and ending at 1 since we're adding up all the probabilities) for each word index.
And when it comes to getting indices for negative sampling, it finds the indices using np.searchsorted which finds the indices the generated random numbers (between 0 and 1) would need to be added to maintain order of the array.

The PR also adds a test that makes sure the new method maintains expected frequency of words (in a simple example).

There are a few advantages for this new approach:

• The saved vocab.dat will be smaller*
  • If using a 10M length unigram table vocab is 314MB
  • If using a 100M length unigram table vocab is over 800MB
  • If using the new approach vocab is 239MB
• Loading a smaller vocab (without unigram table) is a little faster
  • From 0.62s for 10M unigram or 0.72s for 100M unigram table
  • Down to 0.55s with cumulative frequencies (no unigram table)
• Computational cost stays the same**
  • Tested with 100 000 repeats (so per individual time is a lot smaller)
  • When getting 3 at a time
    • Unigram table took: 0.5745s
    • New approach took: 0.2842s
  • When getting 9 at a time
    • Unigram table took: 0.5684s
    • New approach took: 0.4996s
  • When getting 18 at a time
    • Unigram table took: 0.8640s
    • New approach took: 0.6791s
  • When getting 27 at a time
    • Unigram table took: 0.8223s
    • New approach took: 0.9714s

* NOTE: The currently most prevalently used vocab has a unigram table with length 10M, but the defaults are now (for a long time) 100M so if a new unigram table is calculated with no extra input, we get 100M length unigram table.
** NOTE: The number of samples required at a time is dictated by the context vector sizes defined in the config (config.linking.context_vector_sizes). These are (by default) 3, 9, 18, and 27.

[tomolopolis]

Task linked: CU-8696nbm03 Remove use of unigram table