fix: address masking issues identified in #44

[ceholden]

Description

This PR is intended to address issues raised in #44, specifically,

1. We should be masking 0% reflectance as invalid (we had already masked <0) ("zero reflectance should be filtered out as well")
2. We should not mask >100% reflectance as these values are often "valid" but >100% because of unmet assumptions about topography (see, Current issues with VI #44 (comment))
3. We should use -19_999 as a nodata value instead of -9999 because -9999 inside the valid range of normalized differenced indexes (which when scaled ranged from [-10_000, 10_000]. This -19_999 was picked for consistency with the Landsat on-demand vegetation index products (e.g., Landsat NDVI)
4. Use the union of surface reflectance masks for all bands ("Pixel with negative or zero reflectance in any band should have no_data for all VIs")
5. We should avoid numerical computation issues when converting from float64 -> int16, specifically overflows that create a "wrap around" when downcasting.

How I did it

1. Change from masked_outside to masked_less_equal and set <= 0
2. Change from masked_outside to masked_less_equal, which no longer masks >100% reflectance
3. Define fill_value per Index enum member, set this fill value when calculating the index, and write the GeoTIFF with the fill value (-19_999)
4. Union the nodata masks by "or"-ing the nodata mask of all bands
5. See below,

For the issue with wraparound when downcasting, the example in test data was for EVI which can exceed [-10,000, 10,000] because it is not normalized. Consider this example from one of the granules in the test dataset,

# without clipping our large positive number becomes negative, which is misleading
>>> np.float64(192_583.3333).astype(np.int16)
-4025

# if we clip first the value is still a large positive number
>>> np.clip(np.float64(192_583.333), a_min=np.iinfo("int16").min, a_max=np.iinfo("int16").max).astype(np.int16)
32767

How you can test it

• I added a specific unit test for (4), "mask bands by union of value range masks from all bands"
• I added a test for all masking related requirements that I'm aware of and test each one as a separate parameterized test run
• I updated the TIF files we keep for the "expected data" so existing unit tests shoudl work

[chuckwondo]

This looks great! I really like the additional tests. I have only 1 completely ignorable comment that's perhaps a bit pedantic (not pydantic). Feel free to skip it.

[chuckwondo]

Fantastic!