Increase abs_separation default value in tweakreg

[braingram]

The default value for abs_separation in romancal (0.1) matches the previous default value in jwst.

However, spacetelescope/jwst#8445 increased this parameter to 1.0 with mention of the need for this to be 1.5x the tolerance (0.7). However a later PR added a slightly different check spacetelescope/jwst#8476 to verify abs_)separation > sqrt(2) * (abs_)tolerance.

The default values in romancal could likely use some updating (and a similar check added) as I just ran into the same issue fixed by the jwst PR where running roman_elp on an uncal file resulted in RuntimeError: Number of output coordinates exceeded allocation which went away when I set abs_separation to 1.0.

The failed filed were simulated images shared with me by @schlafly:

['r0000101001001001001_01101_0002_WFI10_uncal.asdf' and 'r0000101001001001001_01101_0003_WFI10_uncal.asdf']

[mcara]

Current defaults:

separation = float(default=1.0) # Minimum object separation in arcsec
tolerance = float(default=0.7) # Matching tolerance for xyxymatch in arcsec
abs_separation = float(default=0.1) # Minimum object separation in arcsec when performing absolute astrometry
abs_tolerance = float(default=0.7) # Matching tolerance for xyxymatch in arcsec when performing absolute astrometry

For absolute alignment, the settings are bad, as mentioned by @braingram, resulting in multiple sources being within the tolerance region, while xyxymatch allocates enough memory for just 1 match, resulting in the above error. For relative alignment settings are "better", however, a tolerance of 0.7 arcsec is kind of large, IMHO.

[schlafly]

If you want to suggest a different value, Mihai, please do. In practice we don't use relative alignment and just align to Gaia all of the time, so it is somewhat future-proofing rather than something we actually use.

Practically we expect our ultimate alignment to be ~1 mas, so after tweakreg if we're doing worse than 0.1" for brightish stars something is pretty bad. Before tweakreg alignment errors of a few pixels seem plausible to me.

[mcara]

... Before tweakreg alignment errors of a few pixels seem plausible to me.

=> (abs_)tolerance = (a few pixels) * pixel_scale (and round up to a more round value)