Kernel tables v1

v0.23.0: route :n3 to :a3 on uniform grids, tighten constructor errors

Fixes
- :n3 on a fully uniform grid raised a MethodError (no uniform functor).
  It is now routed to :a3, which is identical on uniform grids and faster;
  :a3 already fell back to :n3 on nonuniform grids, so the two names are
  now interchangeable and the grid picks the implementation.
- ConvolutionInterpolation called directly with a Symbol kernel on a
  uniform grid failed (getindex on Symbol): the uniform builder received
  the raw keyword instead of the normalised kernel tuple. Pre-existing,
  masked by convolution_interpolation always passing a tuple.

Error handling
- FastConvolutionInterpolation now errors on nonuniform grids with a
  message pointing to convolution_interpolation or ConvolutionInterpolation,
  instead of silently interpolating on the wrong spacing.

Docs
- Only :a3 falls back to :n3 on nonuniform grids; :a4, :a5, :a7 are
  uniform-only and error. Docstrings corrected to match the code.

Tests
- Added: :n3/:a3 uniform-grid equality, direct-constructor Symbol kernel,
  nonuniform rejection in FastConvolutionInterpolation, :a4/:a5/:a7
  rejection on nonuniform grids.
- Gaussian kernel test tolerance adjusted.
- 1913 tests passing.

Housekeeping
- Manifest.toml removed from the repo (added to .gitignore).
- SparseArrays compat updated.