Modernise bump() API and add multi-backend support

[brendancol]

Problem

bump() in xrspatial/bump.py is inconsistent with every other function in the library:

1. Wrong input API: Takes raw width/height integers instead of a DataArray template. Every other function takes a DataArray.
2. No backend dispatch: No ArrayTypeFunctionMapping, no cupy/dask backends — numpy only.
3. No coordinates on output: Returns a DataArray without coordinate values, making it incompatible with downstream functions that require coordinate information (slope, proximity, etc.).
4. Performance: The numba-JIT triple-nested loop is O(count × spread²) per bump. Fine for small arrays but slow for large count/spread.

There's even a TODO in the source: "TODO: change parameters to take agg instead of height/width".

Proposed Fix

• Refactor bump() to accept a template DataArray (matching generate_terrain() pattern).
• Add ArrayTypeFunctionMapping dispatch with numpy/cupy/dask/dask+cupy backends.
• For CuPy: @cuda.jit kernel with each thread handling one bump.
• For dask: map_blocks since each bump's effect is local to its spread radius.
• Set proper coordinates on the output DataArray.
• Update README feature matrix.

Breaking Change

The function signature changes from bump(width, height, ...) to accepting a DataArray. Document in changelog.