Convert single-byte charset probers to use nested dicts for language models

[dan-blanchard]

This isolates one of the major changes in #99, which is changing our single-byte charset prober language model format to use nested dicts instead of giant lists and offset math. This makes the code much easier to understand and language model access takes about 60% of the time it used to.

# Current
In [40]: %timeit [lang_model_tuple[(i * 64) + j % 64] for i, j in random_index_tuples]
10000 loops, best of 3: 168 µs per loop

# Same as current but language model is a list instead of a tuple
In [41]: %timeit [lang_model_list[(i * 64) + j % 64] for i, j in random_index_tuples]
10000 loops, best of 3: 170 µs per loop

# Single dictionary, but keys are tuples
In [42]: %timeit [lang_model_tup_dict[(i, j)] for i, j in random_index_tuples]
10000 loops, best of 3: 140 µs per loop

# Nested dictionary like in this PR
In [43]: %timeit [lang_model_nested_dict[i][j] for i, j in random_index_tuples]
10000 loops, best of 3: 99.4 µs per loop

The language model conversion script I've included in this PR does not need to stick around in master long term; I just wanted it here for review, since looking through the code that converts the language models and seeing if that looks right is much easier than visually comparing giant language model files.

I'm still seeing some test failures on this branch where Hungarian is being over-predicted, so this isn't quite ready to merge yet, but I figured putting it up here someone might notice something I missed.

[sigmavirus24]

So, the only performance benefit to using iteritems anywhere is if you have a dictionary with millions of item pairs. If that is where we are (GitHub won't show these diffs) then that's fine, otherwise, I'd rather we just use .items() everywhere. Either way, this looks great. 🎉 🍰:sparkles:

[gimyjendirx]

This isolates one of the major changes in #99, which is changing our single-byte charset prober language model format to use nested dicts instead of giant lists and offset math. This makes the code much easier to understand and language model access takes about 60% of the time it used to.

# Current
In [40]: %timeit [lang_model_tuple[(i * 64) + j % 64] for i, j in random_index_tuples]
10000 loops, best of 3: 168 µs per loop

# Same as current but language model is a list instead of a tuple
In [41]: %timeit [lang_model_list[(i * 64) + j % 64] for i, j in random_index_tuples]
10000 loops, best of 3: 170 µs per loop

# Single dictionary, but keys are tuples
In [42]: %timeit [lang_model_tup_dict[(i, j)] for i, j in random_index_tuples]
10000 loops, best of 3: 140 µs per loop

# Nested dictionary like in this PR
In [43]: %timeit [lang_model_nested_dict[i][j] for i, j in random_index_tuples]
10000 loops, best of 3: 99.4 µs per loop

The language model conversion script I've included in this PR does not need to stick around in master long term; I just wanted it here for review, since looking through the code that converts the language models and seeing if that looks right is much easier than visually comparing giant language model files.

I'm still seeing some test failures on this branch where Hungarian is being over-predicted, so this isn't quite ready to merge yet, but I figured putting it up here someone might notice something I missed.

Marcopolo