Infer range from body of `ForN` loops + packed loops

[bvssvni]

Closes #181

This PR adds support for range inference and packed versions of loops.

For example:

fn randomize_tensor(mut tensor: [[[f64]]]) {
    for i, j, k {
        tensor[i][j][k] = random()
    }
}

Notice that there is no range information in the loop, and 3 indices are specified at once. Such notation is common in mathematics, because this information can be inferred from the body of the loop.

What happens behind the scene is Dyon first expanding the packed loop into 3 loops:

for i {
    for j {
        for k {
            tensor[i][j][k] = random()
        }
    }
}

Then it runs an inference on the body for k, and since neither i or j is declared in that loop, it infers the range to [0, len(tensor[i][j])), by cloning and truncating the AST item tensor[i][j][k] to where it finds the index.

This process is repeated for j, which gets the range [0, len(tensor[i])).

Then again for i, which gets the range [0, len(tensor)).

Now we have the full loops:

for i [0, len(tensor)) {
    for j [0, len(tensor[i])) {
        for k [0, len(tensor[i][j])) {
            tensor[i][j][k] = random()
        }
    }
}

You can also write ranges in the packed version, and depend on previous index:

// For all pairs...
for i n, j [i+1, n) { ... }

The inference and packed versions also works for sum/all/any/max/min/sift loops.

The min and max loops went through a breaking change to make them nicer under composition. See #181 for more information.

[bvssvni]

Updated #116