Fix Simulation.filter_indices bug

[chris-langfield]

This PR has been rewritten to fix a bug in Simulation where filter_indices was initialized as type float, causing some of the confusion that led to the original attempt at patching the metadata.

As discussed with @garrettwrong, casting to object actually does serve to protect the original dtype of the column, mostly as a workaround to pandas-dev/pandas#4094

I actually think the line of attack that this PR originally introduced could make the problem worse, so am tabling it for now.

Original content of the PR:

---

If you try to update a metadata column at only some indices, the entire column's dtype is set to object, erasing whatever dtype the column previously had.

This patch checks the existing dtype of the column and updates the entire series accordingly. Due to a known issue with pandas.DataFrame.update() this will not work (and will be difficult to make work) for int or bool types. See discussion in this thread.

This PR therefore clarifies the following expected behavior for setting metadata dtypes:

• Setting, getting, and updating float32 columns is totally consistent. If the column was added as a float32, it can be updated even with other numeric types, and will still return as float32 values.
• Same for float64
• Int columns will be upcast to float64 if you try to partially update them. The workaround is to not do a partial update, but reset the entire column with the values you want, including ones that won't change.
• Same for bool

These behaviors are tested in test_simulation_metadata and it also tests and has examples of how to do the workaround.

[codecov]

Codecov Report

Merging #816 (f4a2de1) into develop (1187c1b) will not change coverage.
The diff coverage is 100.00%.

@@           Coverage Diff            @@
##           develop     #816   +/-   ##
========================================
  Coverage    88.25%   88.25%           
========================================
  Files          113      113           
  Lines         8891     8891           
========================================
  Hits          7847     7847           
  Misses        1044     1044           

Impacted Files	Coverage Δ
src/aspire/source/image.py	97.47% <100.00%> (ø)

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[chris-langfield]

This may not be possible to implement for integer or bool types.

https://stackoverflow.com/questions/36743563/preserve-dataframe-column-data-type-after-outer-merge

Doing so would involve doing an internal conversion to a Pandas ExtensionDtype

Although this type of int plays nice with DataFrame.update(), this will cause problems down the line when we try to get the metadata by using DataFrame.to_numpy(), which does not know what to do with ExtensionDtypes.

The upshot is there are two options:

• Merge this patch, which preserves dtypes for float32 and float64 columns (maybe important for storing CTF parameters etc based on src.dtype and not accidentally changing precision level). However, in this case Int64 values are upcast to Float64 when you try to do an update(). We would have to remember throughout the codebase to set integer-valued columns all at once (i.e. with values of length n), even if we are just updating some indices.
• Keep the old behavior, where partial updates always cast to object.