All execution places are now modeled as grids:
- Scalar places (host, device) are 1-element grids
- Multi-device grids remain as before

Key changes:
- Added get_dims(), get_place(idx) to exec_place::impl
- Changed activate/deactivate to take index parameter
- Moved set_current_place/unset_current_place to exec_place base
- Deprecated is_grid() in favor of size() > 1
- Updated all client code to use new interface

This eliminates special-casing for grids vs non-grids and allows
uniform iteration over any exec_place.

Made-with: Cursor

The shell class added no value - just use the dynamic interface
directly and return exec_place from the factory methods.

Made-with: Cursor

Same as exec_place_cuda_stream - the shell class adds no value.
Use the dynamic interface directly and return exec_place from factory.

Made-with: Cursor

- Remove meaningless const from return-by-value affine_data_place()
- Remove virtual is_grid() from impl - just use size() > 1 directly

Made-with: Cursor

The condition `is_grid()` was changed to `size() > 1` but this broke
1-element grids (e.g. `all_devices()` on single-GPU systems).

Fix 1: Restore is_grid() to detect any grid by checking if
affine_data_place() is invalid (only grids have invalid affine).

Fix 2: Factory functions now collapse 1-element grids to scalars:
- make_grid(), all_devices(), n_devices(), repeat()
- partition_cyclic(), partition_tile()

This ensures that by construction, any true grid has size() > 1,
making `size() > 1` equivalent to `is_grid()` in practice.

Benefits:
- Simpler mental model: grids always have multiple elements
- No edge cases for 1-element grids
- Single-GPU `all_devices()` returns `device(0)` directly

Return type changes (exec_place_grid -> exec_place):
- all_devices(), n_devices(), repeat(), make_grid()
- partition_cyclic(), partition_tile()
- place_partition::as_grid()

Made-with: Cursor

With 1-element grids now collapsed to scalars by factory functions,
is_grid() is equivalent to size() > 1. Remove the method and use
the simpler size check directly.

Made-with: Cursor

Update data_place::composite() to accept const exec_place& instead of
const exec_place_grid&. This allows passing exec_place from factory
functions that now return exec_place (like repeat(), all_devices()).

Also update:
- data_place_composite to store exec_place instead of exec_place_grid
- get_grid() to return const exec_place&
- localized_array constructor parameter
- slice interface local variables

Made-with: Cursor

exec_place_grid was a shell class that provided no additional value over
exec_place now that all places support the unified grid model. This change:

- Removes exec_place_grid class, keeps impl as exec_place_grid_impl
- Updates make_grid() to return exec_place directly
- Changes as_grid() to return const exec_place& (just returns *this)
- Updates place_partition constructor to take const exec_place&
- Updates loop_dispatch template parameter to exec_place
- Removes deleted parallel_for(exec_place_grid, ...) overload
- Updates C API (stf.cu) to use exec_place* instead of exec_place_grid*
- Removes forward declarations of exec_place_grid

The C API types (stf_exec_place_grid_handle) are unchanged as they are
opaque void* handles that don't depend on the C++ class name.

Made-with: Cursor