Implement on-device general tally/histogram

[sethrj]

For #1751 , #809, #261 we will want some multi-dimensional (potentially ragged) data structures for binning tracks, scoring them and their moments, reducing and normalizing in parallel (threads + MPI processes), and efficiently transferring between CPU and GPU. We might be able to use some other software (ADIOS?) as a backend for reducing and I/O.

Indexers

Continuous:

• One-dimensional nonuniform grid (for x, y, z, E, or t)
• One-dimensional uniform grid (optimization of 1D nonuniform, but common)
• One-dimensional uniform log space (optimization of 1D nonuniform)
• Particle attributes (charged vs neutral, em vs hadron, see https://github.com/celeritas-project/celeritas/issues/1594)

Discrete:

• Particle type
• Volume or region

Multi-dimensional:

• Multi-D regular grid (x * y * z)
• Multi-D ragged (particle type * energy per particle)
• Concatenated grid (e.g. LZ time grid, just an optimization for 1D nonuniform grid)

Index quantity

Continuous:

• Track position
• Track kinetic energy
• Source position (aka primary vertex)
• Angle (either dot product with surface normal or absolute direction?)
• Custom, e.g. (primary position, detector id) -> hypot(detector center - primary position) for LZ (#1751)

Discrete:

• Particle type
• Track's pre-step volume
• (For surface detectors) post-step volume?
• Particle charge

Tallying quantities and estimators:

• Weighted path length (#809)
• Reaction rate (path length estimator)
• Reaction rate (collision estimator)
• Energy deposition
• Change in angle (?)
• Step count (for steps-per-track diagnostic)
• Number of secondaries
• Interacting process (post-step action)
• Loss in energy

Accumulation event:

• Lifetime (for steps-per-track diagnostic, observables)
• Surface crossing (optical detectors, or cell based detectors)
• Virtual surface crossing (mesh tally)
• Every step
• Every interaction

Accumulation type:

• Mean
• Mean + variance
• If we ever do variance reduction schemes, a bin structure that could resolve low-variance but accurately tally high-variance events

Combining dimensions:

• N-dimensional binning: arbitrary number of axes but optimize for a few?
• Ragged binning, e.g. different energy bins for different particle types; perhaps as sub-dimensions (i.e. two concatenated uniform grids would have sub-indices (0, 0..4) and (1, 0..9); the particles could have different sub-dimensions).
• Interface for tallying from Celeritas kernel or using Geant4 tracking+stepping action
• Future work: MSOD (?)

Caching:

• Some indexers (particle type, primary position) will not change over track's lifetime
• Some (energy, volume) will have multiple consecutive accumulation
• Some quantities (change in angle) require both pre- and post-step quantities

[sethrj]

• Quantifier: given particle view (and sometimes params data), return a scalar (or vector? or tuple?)
  • Volume ID: GeoParams -> VolumeId
  • Detector ID: VolumeId -> DetectorId
  • Distance to detector: DetectorId -> real
  • Spatial location: GeoParams -> {real, real, real}
  • Particle type+energy: ParticleParams -> {pid, real}
• Indexer: essentially like Grid, maps value to a bin, has size, composable (can use other indexers)
  • Uniform grid: real -> size
  • Concatenated grids: ? -> size
  • Regular multi-D grids: {size, ...} -> size
  • Irregular grids: {...} -> size
• Estimator: get the quantity to be tallied (maybe same as quantifier?)
  • Unity: increment every time a score happens
  • Weight: increment by a weighted value

Final indexing scheme:

• [permanent index][temporary index]
• permanent index: calculated at track start (source location, particle ID[?], ...)
• temporary index: merged from all indexers

To start:

• Focus on the specific LZ application: two specific tallies
• To start, don't make the indexing fully user configurable/extensible
• Use hard-coded composition to combine indexers into one permanent, one temporary
• Score into double precision, use a single batch