Launch kernels over a list of indices

[giordano]

Part 2/2 of upstreaming Oceananigans piracy of KA (part 1 is #764). Ref: #757. CC @simone-silvestri.

ndrange may now be a device vector of CartesianIndex{N} or
NTuple{N, <:Integer} elements. Each work item handles one listed index:
@index(Global, Cartesian) and @index(Global, NTuple) return it,
@index(Global, Linear) its position in the vector, @ndrange() the
vector's length. The kernel must have a dynamic ndrange and a 1-D
workgroup size, static or given at launch; the partial last workgroup is
bounds-checked against the vector's length.

This is how Oceananigans iterates over the active cells of an immersed
boundary grid. It currently wraps the kernel function, pirates
Base.getproperty on Kernel, Adapt.adapt_structure on NDRange
and CompilerMetadata, partition, expand, and adds one
@device_override __validindex per GPU backend, which the compiled CPU
backend's own ::Any override shadows.

Implementation:

• IndexMap{N} wraps the vector and yields CartesianIndex{N} on
  indexing; it is the mapping of a 1-D NDRange (MappedNDRange).
  expand and linear_index on a MappedNDRange look the index up.
• partition dispatches vectors to mapped_partition, which validates
  the kernel and workgroup size and partitions (length(map),); the
  dynamic-check flag is computed as for any other ndrange.
• Index validity moves into the generic __validindex(ctx, groupidx, idx), which dispatches on the iteration space (I in ndrange for
  regular spaces, position <= length(map) for index maps). The 1-arg
  __validindex(ctx) is now a plain method built on
  KernelInterface.get_group_id/get_local_id, so the POCL override is
  removed; __index_Global_Linear goes through __global_linear in the
  same way. Backends that still override __validindex(ctx) keep
  working for regular ndranges and should forward to the 3-arg method
  to support index maps.
• Adapt.adapt_structure for NDRange, IndexMap and
  CompilerMetadata moves the vector to the device with the kernel
  arguments; CompilerMetadata treats a vector ndrange as
  CartesianIndices((length(v),)).

[vchuravy]

• ndrange may be a device vector of CartesianIndex/NTuple indices, running one work item per
  listed index (kernel(A, ndrange=active_cells)).

I think that's a step too far.

The user could have simply written:

ndrange=length(active_cells) and then done. idx = active_cells(@index Global)

[giordano]

Closing in favour of #772.