Improve flagging and noise estimation

[tskisner]

We have two primary flag bits for shared and detector data:
"invalid" and "processing". The invalid bit means that the
data does not exist or is wrong / corrupted in some way that
makes it useless for any application. The processing bit
means that the data should not "typically" be used for
science results.

This work attempts to change all built-in simulation and
reduction operators to follow these basic guidelines:

1. Operators that load data:

• Boresight pointing marked as invalid if no pointing
  solution exists
• Detector data marked as invalid if it is missing (e.g.
  dropped packets, etc)
• In some cases, there may be data that should not
  routinely be used for science, and these samples should
  be flagged with the processing bit.

2. Operators that simulate telescope observing:

• Unless the operator is also simulating failed pointing
  solutions or readout problems, no samples should be flagged
  as invalid.
• Some samples might be flagged with the processing bit if
  those features are typically known by low-level data
  acquisition systems. Examples might be cooler cycles or
  scan turnarounds, although those samples might also be
  flagged by downstream operators.

3. Operators that simulate detector timestreams:

• Existing invalid detector samples should have their flags
  passed through and remain invalid
• Detector samples flagged with the processing bit should have
  this bit passed through.
• If the simulation requires data inputs (e.g. pointing), and
  that data is flagged as invalid, then the output samples of
  the simulation operator should be flagged as invalid.
• If the simulation uses inputs flagged with the processing
  bit, it is up to the operator whether to use those samples.
• If the simulation is unable to generate some samples, then
  those samples should be marked as invalid since they do not
  contain the full expected detector response
• If the simulation generates poor-quality data for some reason,
  those samples might be marked with the processing flag

4. Operators that process detector timestreams:

• Existing invalid detector samples should have their flags
  passed through and remain invalid
• Detector samples flagged with the processing bit should have
  this bit passed through as well.
• It is up to the operator whether to use samples flagged with
  the processing bit.
• If the operator is modifying the detector data, and fills
  some samples with garbage, zeros, etc, then those samples
  should be marked as invalid.
• If the operator is modifying the detector data, and is unable
  to use / process some data, then those samples should have
  the processing bit set.

This PR also include improvements to flagging in the Offset template and a typo fix in fitting noise models.

[keskitalo]

We need to work on this PR a little more. There needs to be a distinction between samples that should not be used in template fitting and filtering and samples that are excluded from all processing. Point source flags are a good example of usable but excluded samples. We still want the point sources in the final map. I have flagged completely excluded samples such as turnarounds as "invalid" but now they seem to drift into the processing mask.

[tskisner]

Summarizing a discussion from other channels:

Principles

• Invalid data is a "one way street". Any operator which modifies data and fails to modify some samples should mark those samples as invalid.
• Default flagging masks across all operators should be self-consistent and "just work" for unit tests and typical usage.
• Operators that raise flag bits should support taking the mask to raise as a trait, to support custom use cases.
• Other default flag bits will be defined to support common data selection criteria.

Implementation Plan

• In the default flag mask definitions, make these changes:
  • det_mask_sso set to “4”
  • turnaround set to “2” and changed to “shared_mask_turnaround”
  • elnod set to “4” and changed to “shared_mask_elnod”
  • Similar rename of other ground shared masks
  • Build composite mask aliases:
    • shared_mask_nonscience = invalid | unstable | irregular
    • det_mask_nonscience = invalid | processing | sso
• Go through all operators and update defaults to reflect correct mask (usually invalid or nonscience)
• Go through all operators and ensure that failed data modifications are flagged with invalid.

[tskisner]

Ok @keskitalo , this PR is ready for re-review. There is a small issue when running on GPU with hybrid pipeline data movement, but that will likely just be a typo fix somewhere. Will not merge until those tests pass (which are not tested by CI anyway).

[keskitalo]

Mostly small observations.

The one big question is about det_flag_mask being repurposed and whether that is general enough.

[keskitalo]

Looks great!