Clean up sentence generation

[edanaher]

This started out as a nice and clean:

• Pick a pivot letter for each word based on what letters need practice
• Pick letters before and after that letter based on ngram frequences biased somewhat by what letters need practice.

Adding punctuation added a bit of complexity:

• Some punctuation should always be at the beginning or end of a word.
• We now have two sets to choose from, and a parameter to weight punctuation to fudge this.

Numbers further complicate the pictures:

• Some words should be just numbers
• Some numbers should be in words
• Some numbers should be in specific words (1st, 8th, etc.)

This is pretty ugly now. I think it should work to simply assign each symbol a weight (possibly including a class-weight, possibly including user-tunable per-symbol weights), generate the word, and then adjust it to make sure things like punctuation fit, and possibly accumulate numbers to go into a pure-number word later. More thought is required.