Implement device-side event delivery.

Finalizes #184 and includes fix for #285. This PR constitutes a brutal rebase of the `devel/async-integration` branch onto current master: git patch, baby.

Implements asynchronous integration of cells within an `fvm_multicell` instance together with the code to implement efficient device-side event delivery.

Summarized patch notes from the development branch:

* Add four new views in mechanism state:
   * `vec_ci`: cv index to cell index (w.r.t. one lowered multicell instance) map.
   * `vec_t`: current or integration starting time, indexed by cell.
   * `vec_t_to`: integration stopping time, indexed by cell.
   * `vec_dt`: pre-computed delta between `vec_t` and `vec_t_to`, to address performance regression on multicore backend.
* Extend notion of indexed view in `modcc` to encompass `CellIndexedVariable` variables, which have a per-cell value.
* Add/change 'built-in' variables `t`, `t_to`, and `dt` for NMODL.
* Additional utility functions:
   * `util::minmax_value`: return one-pass reduction for minimum and maximum.
   * `util::fill`: range-compatible wrapper for `std::fill`.
   * `util::empty` now works across arbitrary sequences (defaults to calling `empty()` method if it exists).
   * Extends C++17-like functionality for path testing and manipulation (c.f. PR #233).
* Use a cached copy of the per-cell time vector for queries in `fvm_multicell`.
* Modifies back-end matrix assembly, threshold detection and stimulus implementations to be asynchronous-friendly.
* Rename some back-end mechanism variables for clarity, now that there are multiple sorts of cell index vectors present. `cell_index` for example has been renamed to `cell_cv_divs` where it acts as a partition division of cv indices (length 1+#cells), and `cell_to_cv` where it acts as a map from cell to index of first cv for that cell (length #cells).
* Move common event-delivery types (`target_handle`, `deliverable_event`) to `src/backends/event.hpp`.
* Add `multi_event_stream` data structure for `gpu` and `multicore` backends, which handles multiple streams of events, bulk loaded but individually popped.
* Add unit tests for `multi_event_stream`.
* Document event delivery process and abstraction in `src/backends/event_delivery.md`.
* Virtualize `mechanism::deliver_events`.
* Perform event delivery in `fvm_multicell::step_integration()`, using the backend-provided interfaces.
* Add zero dt check/support to matrix state implementations.
* Add zero dt unit tests for multicore and gpu back-end matrices.
* Add debugging helper `util::sepval` for printing/tracing container values.
* Add trace csv output option to miniapp.
* Add GPU kernel and unit test for end-of-integration time step test. By default, keep using the copy-to-host-and-test method, as it is faster for cell counts up to circa 10k; investigate adaptive/threshold solutions in the future.
* Explicitly compute and store per-cell and per-compartment `dt` from integration time bounds.
* Determine lower bound on number of integration steps per interval in order to avoid explicit checking of minimum cell times each step.
* Avoid any time value checking for samplers in the integration loop if none of them could be triggered in the interval.