Fix bug in array_handling.stack_by_init_date

[dougiesquire]

This PR tries to fix a few bugs in array_handling.stack_by_init_date function. I have also changed the API of the function in the following ways:

• remove the freq parameter,
• expect that the time dimension of ds has the same dtype as init_dates. I think this is cleaner than requiring that init_dates is an array of strings,
• Add optional parameters for the names of the time, initial date and lead time dimensions.

I trialled adding a catch to use the stride option in .rolling.contruct when the initial dates are evenly spaced along the time dimension, but this ended up being overly complicated and didn't really add any performance benefits.

I've added some basic tests of the new function and added pytest to environment.yml so that these tests can be run.

Closes #4