Implement out-of-core mapmaking

Using the previously implemented loader pattern, support on demand
loading of observation data when accumulating hits, inverse
covariance, and noise weighted maps for each observation.

- Add PipelineLoader class, which implements a loader which runs
  a Pipeline to generate data for an observation.  This can be
  used to simulate data on demand, one observation at a time.

- Add AccumulateObservation operator, which can accumulate and
  cache or load per-observation map domain products.  This allows
  building RHS map domain objects without keeping all detector
  data in memory, and also allows computing the RHS on smaller
  jobs which are I/O bound.

- Change the BinMap operator to use this new accumulation
  technique.

Use new pixel io interfaces

WIP porting mapmaking utils to new accumulation

Large WIP on per-obs accumulation with optional cache and load.

Author: tskisner

packaging/conda/deps.txt

@@ -4,6 +4,7 @@ libtool
 automake
 psutil
 cython
+ruamel.yaml
 pytest
 tomlkit
 traitlets

setup.py

@@ -325,6 +325,7 @@ def readme():
 conf["entry_points"] = {
     "console_scripts": [
         "toast_env = toast.scripts.toast_env:main",
+        "toast_run = toast.scripts.toast_run:main",
         "toast_fake_focalplane = toast.scripts.toast_fake_focalplane:main",
         "toast_ground_schedule = toast.scripts.toast_ground_schedule:main",
         "toast_project_schedule = toast.scripts.toast_project_schedule:main",

src/toast/CMakeLists.txt

@@ -141,6 +141,7 @@ install(FILES
     observation_dist.py
     observation_data.py
     observation_view.py
+    observation_cache.py
     pointing_utils.py
     vis.py
     rng.py

src/toast/config/__init__.py

@@ -9,6 +9,7 @@
     dump_config,
     load_config,
     parse_config,
+    run_config,
 )
 from .json import dump_json
 from .toml import dump_toml