Add mask_empty_buckets and mask_allnan_buckets methods for BucketResampler

Problem description

After #319 is merged, get_sum will return 0 for empty buckets. Therefore, a method like bucket_resampler.mask_empty_buckets(binned_data, fill_value) is needed to be able to differentiate between empty buckets and buckets with sum 0. This method can leverage self.counts being cached.

Moreover, buckets containing only NaN will also be 0 and have the same problem. Therefore, a method like bucket_resampler.mask_allnan_buckets(binned_data, fill_value) is also needed. Similarly, all-Nan buckets and empty buckets are both NaN in get_average, so the method can be used here as well.

For this to work, a variable like self.nancounts would need to be stored previously in the get_sum method. This is assuming that everything can be done lazily and does not increase the memory consumption. If this is not the case, something like bucket_resampler.mask_allnan_buckets(binned_data, original_data, fill_value) could be implemented.

Expected Output

Being able to differentiate between empty, all-NaN and sum=0 bins.

Actual Result, Traceback if applicable

Not possible at the moment. A workaround using get_count needs to be implemented manually by the user.

Start by reading the BucketResampler entry points get_sum, get_average, and get_count, focusing on the cached counts and handling of NaN values. Define how mask_empty_buckets and mask_allnan_buckets should distinguish empty, all-NaN, and sum=0 buckets, then verify the methods work for both sum and average results without increasing memory use.